YouBank

The AI deal desk that does the work, and learns how you do it.

YouBank is a financial research terminal and an email agent for people whose work is deals. It has three parts:

• Research. A terminal with more than 35 functions and 328 tools built on public regulatory data (SEC EDGAR filings and XBRL facts, Form D, the Treasury, the BLS, public startup directories) plus market prices. Every figure cites its source, and every model shows its working: forecasts with calibrated intervals, credit and earnings-quality models, GARCH risk, Monte Carlo valuation and knowledge tracing that fits the terminal to how well you know it.
• Relationships. An agent that reads your mailbox, drafts replies in your voice, asks you what it cannot answer and remembers the answer, runs cold outreach and nurture, and keeps a pipeline current.
• Autopilot and the adaptive engine. Autopilot sends on your behalf, but only the kinds of email you allow. The adaptive engine decides when a kind of email has earned that, from your own behaviour and with a stated error bound.
• Studio. A live workbook and deck. The agent builds financial models and pitch books while you watch, cell by cell, from SEC data. Slides stay linked to the model, every run is audited and can be undone, and it all exports to Excel and PowerPoint. With the add-in it runs inside Excel and PowerPoint too: the agent writes into your own workbook live, and decks refresh from the model.

Production: https://youbank-nu.vercel.app until the custom domain is set (NEXT_PUBLIC_SITE_URL; see docs/launch-billing.md) · Free while in beta.

Contents

1. Who it is for
2. Features
   • Research terminal
   • Tools
   • Company and filing data
   • Private markets
   • Crypto
   • Fund copies (13F)
   • Algo trading
   • Newsroom
   • AI assistant
   • Studio: live models and decks
   • Relationships: the email agent
   • Autopilot
   • The adaptive engine

---

Who it is for

The product sits between two groups of tools:

• General AI assistants. They will build a DCF for $20 a month, but not from verified data, and they do not do the follow-through. In the September 2026 update of Vals AI's Excel Modeling Benchmark, frontier models passed only 64% of numerical checks.

• Platforms that sell the data. They charge thousands to tens of thousands of dollars a year:

  • Sacra estimates Rogo at about $3,300 per seat a year.
  • Vendr buyer data puts AlphaSense's median contract at $18,375 a year.

  Retail terminals (TIKR, Koyfin) cost $25–$120 a month, but they do not act on anything.

YouBank offers cited, filing-backed work plus an agent that acts on it, priced for the firms in between.

Who	What YouBank does for them
Boutique advisors and placement agents	Comps, buyer lists and teasers without a Capital IQ budget, and an outreach desk that keeps every mandate moving
Emerging VC and PE managers	Form D and funding signals on the companies you track, cited diligence, and a pipeline that files itself from your inbox
Founders raising or selling	Cold outreach, investor and customer follow-up, and replies to your own team, with autopilot where it has been earned
Students recruiting into finance	Real comps and DCFs on real filings, deal walk-throughs, and the same desk they will use on the job

The terminal and tools are tailored to eight careers:

• investment banking
• private equity
• venture capital
• public markets
• corporate finance
• consulting
• accounting
• students

Each career has its own research, tool pack, watchlists, prompt library and default style.

---

Features

Research terminal

/app/terminal is a multi-panel terminal driven from the keyboard, with familiar function codes. Company functions take a ticker (SNOW WACC); market functions do not (ECO); a few take words (EQS profitable software companies growing over 20%, PORT AAPL 40 MSFT 30 KO 30).

Company

Function	What it shows	Models
DES	Description
FA	Financial analysis
GP	Price chart with 50- and 200-day averages, volume, trend and volatility regime, a price cone	GARCH(1,1)
HP	Daily price history
BETA	Beta four ways, with standard errors and a rolling beta	OLS, Blume, Welch (2022)
RISK	Volatility now and forecast, VaR and expected shortfall, drawdowns	EWMA, GARCH, Cornish-Fisher, Kupiec backtest
COMPS	Trading comps
PREC	Precedent transactions
CAP	Capital structure
WACC	Cost of capital from its parts, with a range and a tornado	CAPM, implied-rating spread, Monte Carlo
IRAT	Implied credit rating and default odds	Altman Z'', Ohlson O-score, Merton distance to default
DDIS	Debt maturity ladder and refinancing risk
QUAL	Earnings quality and red flags	Beneish M, Piotroski F, Sloan accruals
FCST	Revenue forecast against the Street	M4 combination, conformal intervals
EE	Earnings: surprises, beat odds, the typical move	Beta-Binomial, conformal quantile
ANR	Analyst ratings, targets and drift
DVD	Dividends and a safety score
FIL / EVT / INS / XBRL	Filings, 8-K events, Form 4 insiders, any XBRL concept
AI	Assistant

Markets and economy

Function	What it shows	Models
WEI / FXC / CMDTY	World indices, currencies, commodities and crypto	Each move in standard deviations of its own volatility
MOST / SECT / MA	Movers, sector performance, M&A filings
GC	Treasury curve against a week, a month and a year ago	Nelson-Siegel; the New York Fed recession probit
ECO	Inflation, jobs, growth and rates, with three-month outlooks	Damped trend, conformal intervals, the Sahm rule
EQS	A screener over every SEC filer, in plain English	A small model turns words into filters you can edit
PORT	Portfolio risk: risk shares, correlations, a year of outcomes	Covariance attribution, bootstrap

Crypto (no ticker needed; free public data, see Crypto)

Function	What it shows	Models
CRYP	The largest tokens, total market cap, bitcoin dominance, DeFi value locked, stablecoin supply
TOKEN <token>	One token: price, supply and float, volatility, drawdown, its protocol's fees, revenue and holders' revenue, valuation multiples, the next unlock	Beta and correlation to bitcoin
DEFI / STBL / YLD	Value locked by chain, category and protocol with fees and revenue; stablecoin supply and net flows; the largest yield pools
BTCN	Bitcoin fees, mempool, hashrate, the next difficulty adjustment, pools	Hashprice, breakeven power price
RAISE / UNLK / TRSY	Crypto venture rounds and lead investors; token unlocks in the next 60 days; companies' crypto at fair value from SEC filings
RWA	Tokenized treasuries, private credit, commodities and real estate
WALLET <address>	Read-only balances and risk of any address (0x…, name.eth, Bitcoin, Solana)	Historical-simulation VaR

Workspace: LEARN (your mastery of each function, what to learn next, quick quizzes), PG (peer groups), TOOLS and TOOL <id>.

What makes it more than a data screen:

• Models and sources on every analytics screen: the method, its assumptions and where the inputs came from.
• AI read: a small, cheap model writes a three-bullet read of any screen from its numbers alone.
• Ask AI hands a screen's question to the assistant, which has the same models as tools.
• Knowledge tracing: typing a command is evidence you know it, clicking a suggestion weaker evidence, closing a panel straight away evidence against. Hints fade as mastery grows and the next functions unlock when their prerequisites are known (Bayesian knowledge tracing with forgetting).
• Fast: every screen but DES and FA loads on demand, data is cached and reused across panels, and hovering a function on the strip prefetches its data.
• Honest about data: a function whose data the market-data plan does not cover says so and points to the ones that work for every US filer.

Type into the command bar, for example SNOW COMPS, CCL IRAT or DDOG TOOL dcf. / or Cmd+K focuses it, ? lists every function. Each seat gets a function strip for its role. A ticker that is also a code works with a function after it: PG DES is Procter & Gamble.

Tools

/app/tools holds 328 tools:

• 230 AI workflows that research and draft;
• 98 calculators that compute exactly, including an inference pack: a Monte Carlo DCF with correlated inputs, a credit scorecard, a forecast with calibrated intervals, a beta estimator, a pipeline forecast, an earnings-quality check and a VaR backtest.

Both kinds render into the same 19 kinds of output block:

• headline numbers, tables, bridges and sensitivity grids;
• risk registers, checklists, timelines and question banks;
• drafted emails.

AI workflows. A workflow's instructions are a numbered method that names the exact data tools and search phrases to use. The model streams tool calls and returns JSON that matches the output contract.

Calculators. A calculator is a pure compute(inputs) function. It recalculates as you type and can be prefilled from the loaded company's SEC data.

Saved runs. Every run is saved with its inputs, output, sources, model and duration. Reopen it from the library at /app/tools/<id>?run=<n>.

How to add tools: docs/05-tool-pack-authoring.md.

Company and filing data

• Fundamentals come from SEC XBRL company facts.
  • LTM is the fiscal year plus year-to-date, less the prior year-to-date.
  • Each figure uses the concept the filer actually used.
  • Restatements are de-duplicated by accession number.
• Filings cover EDGAR submissions and full-text search across every filing and exhibit since 2001. That includes merger proxies, credit agreements, indentures and comment letters, with 8-K item codes decoded.
• Insiders come from Form 4 transactions. Capital structure, precedents, peer groups and saved comps sheets are also available.
• Prices come from Financial Modeling Prep: quotes, market cap and the 52-week range, and on the terminal daily history, estimates, grades, dividends, sectors and M&A filings. The free plan covers daily prices for about 87 sample tickers and caps requests per day; FMP Starter (about $19 a month billed yearly) covers every US symbol, and a display licence is needed before prices are shown to paying users.
• Backups when FMP is out (its daily limit, or a ticker the plan does not cover), in order:
  1. Free sources. Nasdaq's quote API for every US stock and ETF (history, quotes, market cap, dividends, earnings surprises and dates, EPS forecasts, analyst targets and ratings, movers); ETFs that track the indices and commodities it does not list (S&P 500 via SPY, gold via GLD); the ECB's reference rates for currencies; CoinGecko for crypto; the SPDR sector ETFs for sectors; merger filings on SEC EDGAR for deals; and, for dividends outside Nasdaq's listings, dividends per share from each company's own 10-K and 10-Q filings. Dividend history is restated for stock splits, which the filings reveal as a jump in the cover-page share count.
  2. AI research. GPT-5.6 Luna with web search looks up what is still missing (a next earnings date, a consensus figure, a price, a description). A figure is kept only if its page is one the search retrieved, the quote copied from the page contains it, it is recent enough, and it is plausible; when fewer than half survive, one retry goes to GPT-5.4 mini. A lookup costs a few cents, is cached for hours and is logged in the usage ledger. Every screen shows a "Backup data" badge naming the sources used, marks ETF stand-ins ("via SPY"), and lists researched figures with links to their pages. Nasdaq's API is the site's own and is meant for personal use, so the backups suit the free beta; MARKET_BACKUP=off turns off Nasdaq, the ECB rates, CoinGecko and AI research (SEC filings are public domain and stay on). Operators can check which layers answer from production at /api/market/health (Bearer CRON_SECRET).
• Rates and the economy come from the Treasury's daily par yield curve and the BLS (both public domain), which feed the models, and FRED, which is shown with attribution only: FRED's terms bar using its content with machine learning or language models, so it never reaches the assistant or a model.

Private markets

/app/vc holds a directory of 18,037 startups (snapshot of 21 September 2026). Sources:

• Y Combinator, a16z and Thiel Fellows
• Show HN
• SEC Form D
• AI web discovery

It also lists Form D raises with amounts and officers, and crypto projects from DefiLlama's raises database (one entry per project, with every disclosed round, its investors and the total raised).

A nightly job (/api/cron/sync, 06:00 UTC) refreshes the directory:

• Show HN and Form D, for the last three days;
• the YC, a16z and Thiel lists;
• crypto rounds from DefiLlama's free API.

Crypto

/app/crypto (and the terminal's crypto functions) covers crypto research, crypto deals, a read-only portfolio and on-chain records. Nothing in YouBank holds a key, trades, swaps or moves funds. The only thing ever sent to a chain is a transaction the person signs in their own wallet (notarizing), and the server only reads it back.

Research (free). Every figure cites its source under "Models and sources".

• Tokens: prices, market cap, FDV, float, supply and a year of history from CoinGecko's public API (cached minutes; an optional free demo key raises its limits).
• DeFi: value locked by chain, category and protocol, fees, revenue and holders' revenue, stablecoin supply and flows, and pool yields from DefiLlama. A token's page joins the two: market cap and FDV over annualised fees, revenue and holders' revenue.
• Bitcoin: fees, mempool, hashrate, difficulty, pools and mining economics (hashprice, the power price at which a machine breaks even) from mempool.space.
• Tokenized real-world assets: DefiLlama's RWA categories sorted into treasuries (BUIDL, Ondo, Franklin…), private credit (Maple, Centrifuge…), commodities and real estate.

Deal intelligence (free).

• Crypto venture rounds with lead investors, and token unlocks with their share of circulating supply (DefiLlama).
• Public companies' crypto treasuries: crypto assets at fair value from SEC XBRL frames (CryptoAssetFairValue, current and noncurrent, under ASU 2023-08), newest quarter per company.
• Whale and wallet tracking: a watchlist of any public addresses with balances and recent activity.

Portfolio. Connect a wallet (EIP-6963 for MetaMask, Coinbase Wallet and others; the Wallet Standard for Phantom and other Solana wallets) or paste addresses: Ethereum, Base, Arbitrum, Optimism and Polygon (native coins and major tokens, read from public nodes in one multicall per chain), ENS names, Bitcoin and Solana. Holdings are priced and shown with profit and loss against what you say you paid, concentration (Herfindahl), stablecoin share, chain exposure, one-year volatility and one-day 95% value at risk at today's weights (historical simulation; crypto annualises over 365 days).

Notarizing on chain. A deal document, or a Studio model's audit trail (every change, who made it and when, and a hash of the patches), is hashed with SHA-256 in the browser; the file never leaves the machine. The person's own wallet then sends a zero-value transaction to their own address on Base with the hash as readable calldata (YouBank notary v1 sha256:<hash>), costing a fraction of a cent. The server reads the transaction back from a Base node and marks it confirmed only if it is from that address, to itself, carries no value and records that hash. Anyone can check a file against any transaction, and a notarized Studio model can be checked for changes since.

Mining map. Bitcoin mining sites and crypto data centres of the listed miners (Riot, MARA, Core Scientific, Cipher, IREN, Bitdeer, Hut 8, CleanSpark and others), from their filings through mid-2025, with stated capacity, an estimate of annual power use (85% load) and EIA's estimate that crypto mining used 0.6% to 2.3% of U.S. electricity in 2023. Locations are to the town. Sites being converted to AI and HPC are marked. It is a tab on /app/crypto and a layer (with its own button) on the Edge map; the layer is self-contained in src/lib/crypto/map-layer.ts, served as GeoJSON by /api/crypto/sites.

Premium (registered in src/lib/billing/features/crypto.ts; each runs only on a click and the server checks the plan first):

Feature	Plan	What it does	Cost to us per use
Deep wallet analytics	Pro	Every token a wallet holds and its recent transfers and counterparties, from Alchemy (EVM) and Helius (Solana)	about $0.01
Pro crypto data	Pro	CoinGecko Pro and DefiLlama Pro on TOKEN, RAISE and UNLK	about $0.001
Dune queries	Deal Team	The latest results of any saved Dune query, sortable and exportable as CSV	about $0.05

Fund copies (13F)

/app/copy follows hedge funds through their 13F filings and, with Pro, copies them in your own brokerage account. YouBank builds the orders; you approve every batch. Nothing trades without your click, nothing trades from a schedule, and YouBank never holds your money. The full design, data sources, rules and the compliance list are in docs/copy-trading.md.

Timing, said plainly in the product. A 13F arrives once a quarter, up to 45 days after the quarter ends. "Real time" means: the moment a new 13F or 13F/A is filed, we turn it into orders for you to review. Every backtest shows the filing lag.

Free:

• Funds: a curated list (Berkshire Hathaway, Pershing Square, Appaloosa, Baupost, Greenlight, Third Point, Scion, Lone Pine, Tiger Global, Duquesne Family Office and others) and a search of every 13F filer by name or CIK, filled from EDGAR's indexes.
• Fund pages: holdings and weights (a treemap), quarter-over-quarter changes (new, added, trimmed, exited, with stock splits between filings adjusted), concentration, turnover, sectors, top ideas, every filing with its lag, amendments (restatements and new holdings), confidential treatment, and options listed but never copied.
• The clone backtest: buy each filing's top N at the first close after its filing date and hold to the next, against SPY and against an impossible "quarter end, no lag" line. CAGR, volatility, maximum drawdown, Sharpe and beta, with the assumptions and the hypothetical-performance disclosure beside it.
• Paper copies (up to 10) with their record against SPY, rebalanced when the fund files.
• Alerts in the bell (and push) when a followed fund files.

Pro and up (registered in src/lib/billing/features/copytrade.ts; the server checks the plan):

Feature	Plan	What it does	Cost to us
Copy trading with your broker	Pro	Connect about 20 brokers through SnapTrade; review and approve each batch of orders	SnapTrade's charge per connected user, about $2 a month
Fund blends	Deal Team	One copy of several funds, weighted as you choose	none
More live copies	Deal Team	Up to 10 live copies (Pro runs 2)	none

Safety rails. Real orders need COPYTRADE_TRADING_ENABLED=1 (off by default; paper copies work either way). Limit orders by default with a band, per-order and per-day limits, market orders refused while the market is closed, a typed confirmation for large batches, positions you hold outside a copy never sold unless you opt in, consents stored at first use, and an append-only audit log of every consent, proposal, approval and order payload (downloadable as CSV).

Algo trading

/app/algo ("Algo trading" in the nav) runs rule-based strategies on paper or in a person's own broker account, with an AI agent to build, test and explain them. The research (what Webull, Robinhood, SnapTrade, Alpaca, IBKR and Tradier officially allow in 2026, how algo trading is done well, AI-agent guardrails, regulation) and the full design are in docs/algo-trading.md.

• Strategies as data. A typed JSON spec (src/lib/algo/strategy.ts): universe, signal, sizing, schedule, capital, risk limits and execution, built in a form or as JSON. Templates: trend following with volatility targeting, momentum rotation, RSI mean reversion, cointegration pairs, DCA and rebalancing, a 13F copy (reusing Fund copies), covered calls and cash-secured puts. Quant's signals and target portfolios plug in through the contract in src/lib/quant/contracts.ts (src/lib/algo/signals.ts).
• Backtests that say how much to trust them. Event-driven and point in time (fills at the next open), spread, impact and fees, walk-forward folds, bootstrap intervals, the deflated Sharpe with every run counted as a trial, the probability of backtest overfitting, a benchmark, and the hypothetical disclosure.
• Modes. Backtest, paper (the default: simulated fills on real prices), live with approval (each batch tapped) and live autonomous (Deal Team, after 14 days on paper, typed consent per strategy).
• Brokers. The paper broker, Alpaca (OAuth or paper keys), SnapTrade (shared with Fund copies: Webull, Schwab, Fidelity, E*TRADE, Tradier and more), Webull's OpenAPI and Robinhood's Crypto Trading API. Robinhood has no public stock or options API for apps, so those are not offered. Credentials are encrypted.
• The risk engine, on the server, before every order. Per-order, per-day and per-strategy notional, position, gross and net exposure, daily loss stop, drawdown kill switch, order throttles, fat-finger band, no short sales, pattern-day-trader rules, wash-sale warnings, the person's and the administrators' kill switches; an order with an unknown outcome is never resent and blocks new ones until reconciled.
• The agent (Pro): proposes strategies, runs backtests, explains trades from the logged reasoning, tightens limits and pauses things. Orders only from the person's own words, through the risk engine, live ones always waiting for a tap; news and filings are read as data after every tool call has been planned.
• Audit. Every consent, connection, decision, order, block and agent call, append-only, with CSV export.

Feature	Plan	What it does	Cost to us
Live algo trading with approval (algo.live)	Pro	Real orders after each batch is approved	About $0.10 a month of ticks; SnapTrade's fee is Fund copies'
AI trading agent (algo.agent)	Pro	The agent chat	About one assistant answer a turn, inside the AI allowance
Autonomous algo trading (algo.autonomous)	Deal Team	Trades within limits without asking	none beyond the ticks
More strategies and 15-minute decisions (algo.more)	Deal Team	25 strategies (Pro 10, Free 3)	none beyond the ticks

Off by default. Real orders need ALGO_LIVE_ENABLED=1, which should wait for counsel's review (docs/algo-trading.md, section 6).

Newsroom

/app/news is a news desk for each profile: every group of an investment bank (energy gets EIA reports, OPEC, pipeline and LNG approvals; restructuring gets bankruptcies and missed coupons; ECM gets IPO filings and pricings), private equity, venture, public markets, corporate finance, consulting, accounting and students. Every ten minutes it reads about ninety free sources and turns them into stories:

• Publisher feeds. Bloomberg, the WSJ, the FT, NYT DealBook, CNBC, Axios, Semafor, TechCrunch, STAT, the Industry Dive titles, trade press. Headlines and links only; articles are never republished. Open pages (not paywalled, allowed by robots.txt) may be read to write a better summary, but their text is never stored or shown.
• Filing signals from EDGAR's latest-filings feed, often before the press: 8-K items that matter (merger agreements, bankruptcies, restatements, auditor changes, changes of control), IPO filings and pricings, 13D stakes, tender offers, going-private filings, merger proxies, late filings, and Form D raises and fund closes.
• Regulators and wires. Federal Register rules (FERC, FDA, SEC, the Fed, FCC…), the Fed, SEC, FDA, EIA, BLS and BEA, and the PR Newswire, GlobeNewswire and Business Wire press-release feeds.
• A radar for every sector. Early signals before they are news, each linked to its source, a map of where they are happening (bubbles by place, arcs for trade flows and cross-state bank deals), and the sector's deals. Each person picks a sector; the default is their desk's.
  • Technology: the tech radar (Hugging Face daily papers and trending models, GitHub's fastest-rising repositories, Show HN launches).
  • Energy: gas pipelines, LNG terminals and export applications, hydro and nuclear projects moving through FERC, DOE and NRC, at each step; EIA analysis and DOE-funded research.
  • Healthcare: new industry Phase 2 and 3 trials with where they enroll (ClinicalTrials.gov), FDA approvals with new molecular entities and priority reviews marked (openFDA), and Phase 2 and 3 results in NEJM, The Lancet, JAMA, Nature Medicine and JCO (PubMed).
  • Financials: bank mergers, acquisitions and changes in control awaiting the Fed, with comment deadlines (read from Federal Register notices); FDIC failures; Fed, NBER and arXiv research.
  • Industrials and consumer: ITC anti-dumping and patent import cases with the countries involved, Entity List changes and USTR actions; EPA, DOT, OSHA, FTC and CPSC rules; CPSC recalls with units, where sold and where made; robotics research.
  • Media and telecom, real estate: FCC, HUD and FHFA rules in motion; networking and housing research. Built from public sources, cached a day and refreshed by the Newsroom pass; a source that does not answer keeps its last good lane.
• Crypto. CoinDesk, The Block, Decrypt and Blockworks headlines; DefiLlama rounds of $5M and up and token unlocks that add 1% or more to circulating supply; and 8-Ks about bitcoin or digital-asset treasuries from EDGAR full-text search. All tagged with the crypto lens, so NI CRYPTO finds them.
• Research briefs. Twice a day, a small model with web search looks for each active desk's stories that have no feed (Reuters, AP); a story is kept only if its page was retrieved and is recent.

The same event from many outlets becomes one story (headline embeddings plus word and figure overlap, calibrated on live headlines), ranked for each person by desk fit, importance, watchlist, the companies where their contacts work, follows, mutes and freshness, with the reason shown. A small model reads the important stories once: bullets, key numbers, why it matters, companies (tickers checked against SEC's list) and deal terms. Deals and raises fill a tracker with the premium to the unaffected close and implied multiples from SEC figures, and advisor league tables.

• Five editions, chosen in the header or in Settings: Front page (a data-led front: the hero beside its chart, generated data-art thumbnails, "the day in data" drawn in as you scroll, a ticker and the globe), Terminal (a Bloomberg-style wire), Editorial (an FT-style magazine), Brief (an Axios-style morning brief with the wire alongside) and Modern (an Apple News-style dashboard). With the advanced switch, any look goes with any layout. Stories open in a side peek or on their own page; motion is rich, subtle or off (and off under reduced motion).
• A front page that learns, explainably: what you read, save, follow and hide becomes an affinity per desk tag, company and kind of story, fading over two weeks, with the top kept varied. Every card says why it is there, and "Why you're seeing this" shows the whole sum.
• Inside a story: an interactive chart (the price reaction with the moment it broke, the deal on a log scale, the key figures, or how coverage spread), a timeline, a relationship map of who is involved (with your own contacts), and "why this matters to you": your watchlist, contacts, pipeline and Edge watches it touches (free; an AI note on request with Pro).
• The globe: where the last two days' news is happening, markers sized by significance, filtered by desk, one click to the stories.
• Listen, swipe, watch: an audio briefing with a chapter per story (free in your browser's voice; with Pro, an AI script in a neural voice, on a click or each morning), Brief mode (a 20-second card per story, swipe on a phone, keys on a desktop), and a 60-second recap of animated slides you can save as images or record to video in the browser.
• Follow a story to get a bell alert and a push when it develops (five free, unlimited with Pro).
• Public story pages at /news/<story>: the summary, chart, timeline and sources for anyone, with link previews drawn from the story's data, share buttons and a sign-up invitation; never anything personal. In the sitemap.
• The morning brief, written once a day per desk with a "for you" section, on Home, in the app, and by email (from your own connected mailbox), browser push or Slack.
• Alerts for watchlist companies (urgent for a bankruptcy, restatement or takeover), companies where your contacts work (also raised in Relationships as a reason to reconnect, drafted on approval), $1B+ deals and top stories for your desk, with quiet hours.
• In the terminal: TOP (your desk's top stories), CN (company news), NI <topic>, and the latest news on DES.
• The AI budget is capped at $25 a month (), in tiers: personal notes pause first, then research, then briefs; past the cap the Newsroom still works without a model.

AI assistant

One interface over two providers:

• OpenAI: the Responses API, default model gpt-6-astra.
• Anthropic: the Messages API, with extended thinking.

You can switch the model and the reasoning depth (low, medium, high or xhigh) per run. The assistant streams its output, calls function tools, shows reasoning summaries, searches the web and returns structured output.

Thirty-two tools:

Tool	Tool	Tool
get_company_financials	get_trading_comps	search_companies
search_filing	read_filing	web_research
form_d_search	search_startups	edgar_fulltext_search
read_document	get_recent_filings	get_xbrl_series
get_insider_transactions	calc (exact arithmetic)	get_price_risk
get_credit_risk	get_earnings_quality	get_revenue_forecast
get_cost_of_capital	get_macro_outlook	screen_companies
get_crypto_market	get_token	get_defi_overview
get_stablecoins	get_crypto_raises	get_token_unlocks
get_crypto_treasuries	get_bitcoin_network	get_rwa_tokenization
get_wallet_portfolio	token_valuation

The seven after calc are the terminal's models, so an answer about risk, credit, a forecast or the economy names the model behind each number and gives a range. The crypto tools read free public data (CoinGecko, DefiLlama, mempool.space, SEC XBRL, public nodes) and return source ids for every figure; token_valuation does fee-multiple, discounted token cash flow and staking-yield arithmetic exactly.

Cost. Classification, extraction and summaries route to a small model (GPT-5.6 Luna or Claude Haiku 4.5) and drafting to a mid-sized one, which costs a fraction of the flagship; system prompts are kept stable so the providers' prompt caches apply; read-only tools run in parallel; and every call's tokens and cost are recorded and shown in Settings. See docs/research/ai-inference.md.

Allowances. Each person's model spend is capped per UTC day and per allowance month by their plan's AI allowance, from Free ($0.75 a day, $3 a month) to Enterprise ($30 a day, $130 a month per seat). The allowance month starts on the billing anniversary for subscribers and on the 1st otherwise; AI credit packs top it up. Administrators have no cap. Settings, under Plan, shows what has been used. At the cap, AI pauses and everything else keeps working. Prices and allowances: Plans and billing.

Modes: analyst, research, draft, critique, and coach (mock interviews).

Deep research (Pro). Tick "Deep research" under the chat box for a question that matters. The run starts with a short plan, uses up to 30 tool steps (against 10) at maximum reasoning effort on the flagship model (AI_DEEP_MODEL overrides it), checks each figure that drives the conclusion in a second source, and ends with a "Not verified" list. It runs only when ticked; a plan without it gets the plan message in the chat.

Studio: live models and decks

/app/studio is a spreadsheet and slide workspace where the agent does analyst work while you watch. You can start in three ways:

• describe what you need;
• pick a template and a ticker;
• drop in an .xlsx, .xlsm or .csv.

The terminal's company page also links straight to a new DCF, valuation pack, LBO or comps model for that ticker.

Watching the agent work

• The agent's edits stream to your screen the moment each one is stored. New cells appear in reading order and glow, and the agent's cursor shows where it is writing. The view follows it across sheets and slides.
• Other people with the document open see the same edits over a live event stream.
• Every run ends with:
  • an audit of the model, shown as a health strip;
  • a refresh of the data tables;
  • a short summary with cell references.
• Speed: choose Fast, Balanced or Thorough, or Deep (Pro): maximum reasoning, up to 40 steps, then a review pass in which the agent is handed the audit's errors and fixes them before the run ends.
• Stop: halts the agent between steps.
• Undo this run: reverses everything the run did. Every single change, by the agent or a person, is in the History tab with its own undo.

The spreadsheet

The grid is built to feel like Excel:

• a formula bar and name box, selection by click, shift-click or drag, and Excel keyboard navigation;
• on a touch screen: tap a cell to select it and tap it again (or double-tap) to edit in a bar docked above the keyboard, with formula keys and arrows that save and step to the next cell without the keyboard closing; drag the dot at the selection's corner to cover a range;
• type-to-edit and F2, clearing with Delete, and undo with Cmd/Ctrl+Z;
• copy and paste as tab-separated text, to and from Excel. Pasting within Studio adjusts relative references.
• frozen panes, merged cells, text that spills into empty neighbours, and resizable columns;
• number formats and banker colour shortcuts;
• cell comments, source markers, and a "cell types" legend that tints inputs, formulas, links and numbers typed into formulas.

The calculation engine (src/lib/studio/) is YouBank's own, and it runs in the browser and on the server:

• Formula language. A tokenizer and parser follow Excel's rules: negation binds tighter than ^, and ^ is left-associative. They handle sheet-qualified references, whole columns and rows, defined names, error literals and array arithmetic.

• About 120 functions, including:

  • NPV, XNPV, IRR, XIRR, MIRR, PMT, IPMT, PPMT, PV, FV, NPER, RATE, RRI and EFFECT;
  • INDEX, MATCH, XLOOKUP, VLOOKUP, OFFSET and INDIRECT;
  • SUMIFS, COUNTIFS, AVERAGEIFS and SUMPRODUCT;
  • EDATE, EOMONTH and YEARFRAC;
  • TEXT, the statistical functions and the logical functions.

  They are tested against the results Excel documents.

• Recalculation.

  • Values are computed lazily and cached.
  • A dependency graph means an edit recalculates only the cells that depend on it.
  • Deep chains are evaluated bottom-up, so a 6,000-step chain does not exhaust the stack.

• Circular references. They are solved the way Excel's iterative calculation does it. Tarjan's algorithm finds each circular group, which is then solved by Gauss–Seidel substitution, ordered so a loop that is switched off settles at once. The limit is 100 iterations, stopping when nothing moves by more than 0.001. The LBO template charges interest on average debt balances, with a circuit-breaker switch.

• What-if tools. Two-way data tables (the output recomputed for every pair of inputs) and goal seek.

• Imported workbooks. A cell using a function YouBank does not implement shows the value the file had saved.

Templates

Each template is fully formula-driven and filled from SEC XBRL company facts and market data:

• blue inputs, black formulas, green links;
• USD millions;
• key outputs saved as named ranges.

Template	What it builds
DCF	Unlevered free cash flow over 5 years; a CAPM WACC build; mid-year convention; perpetuity-growth terminal value; implied price and upside; a live WACC × terminal-growth grid written as formulas, with its steps as inputs. For a loss-making company, the margin ramps to its margin before stock-based compensation
Trading comps	Peers from the saved peer group or a list you give; EV/revenue and EV/EBITDA with quartiles; the implied value range of the target
Valuation pack	DCF, comps and a summary sheet that feeds a football field, plus a linked 6-slide pitch deck
LBO	Sources and uses; a debt schedule with a cash sweep and interest on average balances; IRR and MOIC; an entry × exit IRR data table. It uses adjusted EBITDA, with a take-private entry at a 25% premium, for listed companies
Merger model	Accretion/dilution, breakeven synergies, and a live premium × stock-mix grid
Cap table	A priced round, with the option-pool top-up solved in closed form
Token multiples	A token's market cap and FDV over its annualised fees, revenue and holders' revenue, with the implied price from low, mid and high peer multiples
Token DCF	Cash flow to tokenholders over 5 years with fading growth, a terminal value, and value per token on the supply expected after unlocks; a discount rate × terminal growth data table
Staking yield	Nominal and real staking yield, the dilution of a holder who does not stake, and the value of fees to stakers; a staking ratio × issuance data table
Crypto comps	Peer tokens (named, or the top fee earners in the token's category) on market cap and FDV over fees, revenue, holders' revenue and value locked, with quartiles and the implied price

The crypto templates take a token symbol or CoinGecko id in the ticker box and are filled from CoinGecko and DefiLlama (USD millions, supplies in millions of tokens); left blank, they open with illustrative inputs.

The deck

• Slides are 16:9. Their tables, charts, figures and football fields are links into the model, so a changed assumption flows to every page.
• Charts come in column, bar, stacked, line, pie, waterfall and football-field forms.
• You can edit titles and text, drag elements to move or resize them (with a mouse or a finger), reorder slides and comment on them. On a phone you swipe between slides, with the thumbnails in a strip below.
• Tie-out checks the deck against the model. It flags:
  • broken links and error values;
  • tables typed in by hand;
  • any number in slide text that is not in the model at the precision shown.

Grunt-work tools

These are available to the agent and one click away in the toolbar:

• Audit: errors, numbers typed into formulas, overwritten formulas in projection rows, formulas that break their row's pattern, references to empty cells, unused inputs, circular references and failed checks. Each finding links to its cell.
• Banker formatting: colours, and number formats chosen by what each row is about.
• Refresh data tables.
• Turn the comments: the agent makes each requested change across the model and the deck, then resolves each comment with a note.
• Intake report for uploads: sheets, cells and formulas; hidden sheets; links to other workbooks; unsupported functions; and what could not be kept (macros, images, conditional formats, charts). A "Review this file" button is included, and file contents are treated as data, never as instructions.
• Uncertainty tools:
  • monte_carlo draws the model's own input cells (optionally correlated through a Gaussian copula), recomputes the whole workbook for each draw without changing it, and writes percentiles, a tornado and each input's share of the variance to a Monte Carlo sheet slides can link to;
  • suggest_assumptions gives data-driven ranges for growth, margin, WACC and terminal growth;
  • forecast_series extends history with calibrated intervals;
  • risk_check adds the implied rating and earnings-quality flags.
• Per-career quick tasks:
  • bankers get valuation packs and football fields;
  • PE gets LBOs, returns attribution and IC decks;
  • VCs get rounds, cap tables and liquidation waterfalls;
  • students get "plant three errors for me to find".

Files

Direction	What happens
Export .xlsx	Formulas with their computed results, styles, widths, frozen panes, merges and defined names. Newer functions get Excel's _xlfn. prefix, and the file asks Excel to recalculate on open. Data tables are written as values with a note
Export .pptx	Native, editable tables and column, bar, line and pie charts. Football fields and waterfalls are drawn as precise shapes, the way think-cell does it. Sources, page numbers and confidentiality lines are included
Import	.xlsx and .xlsm, keeping values, formulas (shared formulas expanded), number formats, fonts, fills, borders, widths, frozen panes, merges and defined names. Also .csv

Inside Excel and PowerPoint

The YouBank add-in puts Studio inside Excel and PowerPoint, on the web, Windows and Mac. Install it from /app/office (the manifest is served at /office/manifest.xml), click YouBank on the Home tab, and connect with a code.

Connecting. Office runs add-ins in a frame where a site's cookies are often blocked, so the add-in signs in with a device code, not a session:

1. The add-in asks for a code and shows it (ABCD-2345, with no look-alike characters).
2. The person approves it at /office/connect, signed in to YouBank. That page sits outside the app, so a first-time user can sign in and approve before the ten-minute code expires.
3. The add-in holds a poll secret that only it knows, and uses it to collect a device token, once.

Only hashes of the poll secret and the token are stored. Every connected install is listed at /app/office, and disconnecting one stops its token at once. The Studio API accepts either the browser session or the token, so the add-in uses the same routes as the browser.

Excel. A workbook links to a Studio document in one of three ways:

• build a model here: a DCF, comps, an LBO or a valuation pack from SEC data;
• use the workbook as it is: a blank sheet or your own model;
• open an existing YouBank model into it.

Once linked:

• The agent writes into your workbook live. Each patch the agent commits streams to the add-in and is applied at once through Office.js.
  • Cells are grouped into runs along a row, one range per run.
  • The selection follows the agent unless you turn Follow off.
  • Everything that writes to Excel goes through one queue, so edits land in order.
  • A formula Excel rejects fails alone; the rest of the change still lands.
• Your edits sync back. A change marks its sheet. A second later that sheet is read and sent, and Studio stores the difference as one undoable change ("Synced from Excel: 3 cells").
  • A snapshot is compared by value, so a recalculated result, a formula's letter case or a source tag does not count as an edit.
  • Renaming a sheet in Excel renames it in Studio, so the deck's links survive.
  • New, deleted and reordered sheets are picked up on a four-second beat.
• A sync cannot undo newer work. The add-in says which change it last applied. If anything newer has landed, the server refuses the sync (409); the add-in catches up and tries again.
• Reopening a workbook days later. The workbook remembers the last change it saw. When it opens, the add-in:
  1. rebuilds that state from the undo patches of everything since;
  2. replays those changes into Excel;
  3. sends anything typed into Excel while it was closed.
• What makes the round trip: values, formulas, number formats, fonts, fills, borders, alignment, column widths, frozen panes, sheet order and defined names. Text that Excel would turn into a number or a date ("2025", "Jun-25") is written with a leading apostrophe, so it stays text.
• Large workbooks. Snapshots are gzipped. A sheet is read up to 150,000 cells, and its formats up to 25,000 cells. A workbook is limited to 400,000 cells.

PowerPoint.

• Insert the deck. The deck arrives after the selected slide, as native tables and charts. Each slide is tagged with its Studio slide and document.
• Refresh from the model.
  • Each tagged slide is replaced in place with its current version.
  • Slides new in Studio are added after the last one; slides deleted in Studio are removed.
  • Slides you made yourself are never touched.
  • PptxGenJS numbers exported slides from 256, which is how one slide is picked out of the file.
• Keep my edits unlinks the selected slides, so a refresh leaves them alone.
• The brand check, the tie-out and the agent are in the pane. After an agent run, the slides refresh.

Requirements. ExcelApi 1.9 and PowerPointApi 1.3: Microsoft 365, Office 2021 or later, or Office on the web.

Checkpoints

• Save the model and deck under a name ("Sent to MD"). You can do it from Studio's History tab, from the add-in, or by asking the agent.
• What changed since is a semantic diff, not a cell dump. It shows:
  • the key outputs (named outputs, key-output cells and slide metrics), with before, after and the change in percent;
  • every input that changed, with its row label;
  • the formulas that changed;
  • sheets and slides added, removed or renamed.
• Two checkpoints can also be compared with each other.
• Restore returns to a checkpoint as one change, which can itself be undone.
• Comparisons ignore key order, because Postgres jsonb re-sorts object keys.

Brand check

Run it from Checks, from PowerPoint, or by asking the agent. It flags what gets a book sent back:

• a cover dated in a past month;
• a page of figures without a source line;
• elements running off the page or into the footer, and elements overlapping;
• text that will not fit its box, text under 9 pt, and colours outside the deck's palette;
• tables with too many rows for their space;
• headline figures without a caption, empty slides and two-line titles.

Most findings come with a fix:

• bring the cover date up to this month;
• add a source line naming the model's sheets;
• move an element inside the page;
• reduce a font size;
• use the theme colour.

Fix all applies every fix as one change. Fixes to the same element stack rather than overwrite each other.

Paper in: marked-up printouts and data rooms

• A marked-up printout becomes comments.
  • Upload a PDF or a phone photo of the reviewer's pen marks.
  • The model reads the handwriting, circles and arrows against an outline of the live document: slides by number, rows by label, periods by column header.
  • Each mark becomes a comment on its slide or cell. Marks it cannot place land on their page's slide, and are counted.
  • Then Turn the comments has the agent make the changes.
• A data-room PDF becomes a sheet of sourced inputs. Upload a CIM, audited accounts or a management deck, and its financial tables become a new sheet:
  • each table sits under a banded title with its unit and page;
  • figures are copied exactly as printed, as blue inputs; parentheses become negatives and percentages become decimals;
  • each figure is tagged with its file and page;
  • notes on restatements, adjustments and unaudited periods are kept at the foot.
• Files can be up to 4 MB: PDF, PNG, JPEG, WebP or GIF. Printed text is treated as content, never as an instruction.

Research behind Studio: docs/research/studio-competitive-landscape.md and docs/research/analyst-grunt-work.md.

Relationships: the email agent

/app/crm has eight tabs: Queue, Inbox, Pipeline, Contacts, Insights, Campaigns, Nurture, and Agent & autopilot.

Mailboxes

A mailbox connects in one of two ways:

• App password (IMAP and SMTP). Presets cover Gmail and Google Workspace, iCloud, Yahoo, Zoho and Fastmail.
  • Mail is read over IMAP (imapflow) and sent over SMTP (nodemailer), from your own address.
  • This is the fastest route and needs no Google review.
• Gmail API with OAuth. Available when GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET are set.
  • It uses the gmail.readonly and gmail.send scopes only.
  • Sync is incremental through Gmail's history API.

More than one mailbox (a work and a personal address, or a shared deal inbox) is part of the Deal Team plan; the first mailbox is open to everyone. The check runs before a second address is stored.

Threading.

• Threads are keyed by Gmail's thread id where the server exposes it (X-GM-THRID). Otherwise the key is the root Message-ID, found from References and In-Reply-To.
• Replies carry In-Reply-To and References, so they thread in the recipient's mail client.
• A reply is sent from the mailbox that received the thread.

Two modes

The mode changes the triage prompts, the pipeline and the writing role.

Mode	For	Pipeline
Deal flow	Investing in or advising companies	inbox → screening → diligence → partner review → term sheet → portfolio (or passed)
Business development	Selling, partnering, fundraising	lead → contacted → engaged → meeting → proposal → won (or lost)

Triage

Every new thread is classified. Triage records:

• A category. One of: prospect, customer, partner, colleague, founder pitch, intro request, portfolio update, LP or investor, scheduling, diligence material, vendor, recruiting, newsletter, other.
• An audience. A thread is internal if the sender is at one of your company's domains, and external otherwise.
• Whether it needs a reply, decided by deterministic rules together with the model.
• The stage the email implies for the deal.
• Whether the sender opted out. "Please remove me" is an opt-out; "not right now, try us next year" is not.

Newsletters, auto-replies, no-reply senders and bulk mail are recognised and never answered.

Drafting

Every draft is written as you. The draft prompt includes:

• your "About you" and signature;
• your standing instructions and knowledge (pasted, or uploaded as .txt, .md or .csv);
• a writing voice learned from your sent mail (a description of how you write, with no details about the people you wrote to);
• your playbook answers and active lessons;
• the three most similar emails you have sent yourself, as examples of tone.

Each draft carries:

• a rationale;
• a confidence (high, medium or low);
• a sensitive flag for money, terms and legal matters. Quoting a price exactly as you authorised it is not sensitive.
• the questions it needs you to answer.

It asks, and remembers

When a reply needs something only you know, such as a price, a date or a yes, the agent asks one short question in the queue.

• Your answer completes the draft.
• Tick "remember" and the answer goes into the playbook. The model first rewrites it so that it answers the next person who asks, not only this thread.
• Every later draft uses the playbook, so the same question is not asked twice.

The review queue

Everything the agent writes appears in one place:

• questions waiting for you;
• offers to put a kind of email on autopilot;
• suggested next steps;
• a composer for new emails;
• scheduled automatic sends, each with a countdown and a Stop button;
• drafts waiting for review;
• recently sent mail.

You can edit a draft before sending it, and it goes out from your mailbox.

Suggestions

The agent proposes next steps, each with its reasoning:

• a stage move the latest email implies;
• a follow-up when nobody has answered (after 5 days by default);
• a check-in on a deal that has gone quiet (21 days by default);
• a reconnection when a contact's company files a Form D. Funding signals are matched to your contacts by company.

Nurture

Nurture rules find relationships that have gone quiet. For each one, the agent decides whether there is an honest reason to write. When there is not, it logs why. Each rule has a daily draft cap.

Campaigns

Campaigns are part of the Deal Team plan: creating, launching, qualifying and drafting them, by hand or on the heartbeat and the nightly run (both skip campaigns for plans without them). Pausing, editing and archiving stay open, so nothing is stranded when a plan changes.

Cold outreach, from lead list to follow-ups:

• Leads come from the startup directory (filtered) or from a pasted list. Pasted leads without an email address are counted as skipped, not dropped silently.
• Qualification checks each lead against an ideal profile before anything is written.
• Sequences are personalised and multi-step. A sequence stops when the lead replies or opts out.
• Consent guard. Recipients at EU, EEA and Canadian domains are skipped unless you confirm that you have consent or another lawful basis.
• Cross-account cap. No recipient gets more than one first cold email through YouBank in 30 days, counted across every account.
• Experiments. The first email's opening angle and send time are chosen by the adaptive engine's outreach experiments.

Contacts, pipeline and compose

• Contacts are grouped by kind: prospect, customer, partner, colleague, founder, investor, LP, banker or operator.
• Deals move through the mode's stages.
• Threads show their activity, including what autopilot did.
• Compose writes a new email from a one-line instruction.

Insights

Relationship intelligence from the mail the agent has read, each number with its method:

• Relationship strength (0 to 100): recency-weighted emails times how two-way they are, scaled to your own volume.
• Reply odds for a new email to each contact: a logistic model trained on your own settled emails, used only when it beats per-contact reply rates on your most recent quarter (Brier score shown).
• Waiting on a reply: Kaplan-Meier over every email says when an unprompted reply becomes unlikely; threads past that point are marked "nudge due", with a one-click nudge draft.
• Going quiet: people who wrote regularly and are well past their usual gap, with a reconnect draft.
• Pipeline forecast: each open deal's win odds (stage priors calibrated to your record, moved by engagement), the exact distribution of wins and a simulated P10/P50/P90 value.
• What each contact knows: every email is tagged once with its topics by a small model; replaying a contact's mail gives, per topic, the chance they know it (knowledge tracing with forgetting) and how much they engage with it. Drafts to them are told not to re-explain what they know and what is worth raising next.

The morning run

/api/cron/agent runs daily at 11:00 UTC. It works in this order:

1. The scans, which only read data: funding signals, follow-ups and stale deals.
2. The work that calls the model: nurture rules and due campaign steps. Each has its own cap, and the whole run has a time budget.
3. Settling the outreach experiments.

It only drafts. Drafts of a kind that is on autopilot are scheduled, and the send queue sends them within your sending hours. Each person can turn the morning run off.

Autopilot

Autopilot is off by default. When it is on, each kind of email has its own setting: Off, Ask me or Autopilot.

Sending on its own is part of the Deal Team plan. Switching it on checks the plan, and the scheduler checks it again before every automatic send: if a plan no longer includes it, the master switch counts as off and anything already scheduled waits in the review queue with the reason. Reading, triage and drafts that wait for you stay open to every plan.

Kind of email	Default
Replies to people outside your company	Ask me
Replies to your coworkers	Ask me
Cold email campaigns	Ask me
Follow-ups when nobody answers	Off
Reconnecting with quiet contacts	Ask me

Other settings:

Setting	Default	What it does
Hold before sending	5 minutes	A window in which to stop any automatic email
Daily cap	40	Maximum automatic sends per rolling 24 hours, across everything
Sending hours	08:00–19:00	In your time zone
Weekdays only	On	No automatic sends at weekends
Read my mailbox every few minutes	On	Lets the heartbeat sync the mailbox
Company email domains	Your mailbox's domain, unless it is a public provider such as gmail.com	Who counts as a coworker

What autopilot will not send

Checks run when a draft is considered, and again just before it is sent. A draft that fails any check comes to you with the reasons written down.

• The agent's confidence is not high.
• It touches money, terms, legal or anything sensitive.
• It needs an answer or a decision from you.
• It contains a blank to fill in.
• It is empty or has no recipient.
• The recipient is an automated address, or the draft answers an automated message.
• The recipient asked not to be contacted (for anything that is not a reply to them).
• Autopilot has already sent twice in this conversation today.
• Security veto: it contains a link, email address or account-like number (8+ digits, or IBAN-shaped) that does not appear in your own settings, knowledge, playbook or signature. An inbound email therefore cannot steer the agent into sending an attacker's link.
• Earned autonomy: that kind of email is on probation, or this draft was picked as a spot check (see the adaptive engine).

How a send happens

1. A scheduled draft waits out the hold, and then for your sending hours.
2. Just before sending, the send queue re-reads the live thread. If you already replied (from your phone, say), or a newer message changed the question, the draft is withdrawn.
3. The send claims the draft (pending → sending), so a draft can go only once. A send stuck in sending is recovered from the time it was claimed.
4. sendDraft() in src/lib/crm/send.ts is the only code that sends. It adds your signature and threading headers, refuses an opted-out recipient, and records who sent it (you or autopilot).

The heartbeat

Vercel's Hobby plan runs cron at most once a day, which is too slow for an agent that answers email. Instead:

• A Neon Function (neon/autopilot.ts) with a schedule trigger (0,20,40 * * * *) calls /api/cron/autopilot with a bearer secret; neon/news.ts calls /api/cron/news a minute later. Every twenty minutes, so the free plan's compute can sleep between passes. On Supabase, scripts/db-setup.mts finish creates pg_cron jobs instead.
• The function relays only genuine trigger deliveries. Neon strips any x-neon-* header a caller sends, so an x-neon-trigger-invocation-id that matches the body's invocation_id can only come from the trigger.

Each pass, for each person with autopilot on:

1. Take a per-person lock, so passes never overlap. A manual sync takes the same lock.
2. Sync the mailboxes. Then triage and draft each new thread, and consider each draft for sending.
3. Draft due follow-ups, unless follow-ups are off.
4. Prepare due campaign steps.
5. Send whatever is due.

Regulated mode

Regulated mode is for FINRA and SEC registrants.

• Autopilot cannot be switched on.
• Every email is sent by a person, from their own mailbox, so the firm's archive captures it.
• An audit log (CSV) lists every email the agent drafted, who sent it, and how much it was edited (/api/crm/audit).

The adaptive engine

The principle. A language model's confidence in itself is not evidence. It stays a hard gate: nothing below high confidence is sent automatically. But what to automate is decided from what the person does.

The code:

• src/lib/crm/engine-rules.ts: the pure rules. The public site's live demo runs the same file in the browser.
• src/lib/crm/engine.ts: the server side.
• src/lib/crm/learning-math.ts: dependency-free maths:
  • the Beta CDF (Lentz's continued fraction) and its quantile (bisection);
  • Lanczos log-gamma;
  • Marsaglia–Tsang gamma sampling, for Beta sampling;
  • a seeded mulberry32 generator.

There are three learning loops per account.

Loop 1: earned autonomy

Labels. Every draft the person sends is a label.

• The edit ratio is the word-level Levenshtein distance between draft and sent text, divided by the length of the longer one. The signature is stripped first.
• A draft is good if the ratio is at most 0.05 and no critical field changed: a number, amount, percentage, link, email address, weekday or month.
• Otherwise it is bad: edited (ratio ≤ 0.25), rewritten, or a critical change. A light edit still counts as bad, because autopilot would have sent it wrong.
• Discarded drafts are not labels, because "no reply needed" and "wrong reply" look the same from outside.
• Stopping an automatic send counts as a bad label and extends the cancel streak.

Strata. Trust is kept separately for each kind of email, audience and category. Examples: reply:internal:colleague, reply:external:prospect, campaign:first, campaign:follow, follow_up, nurture and compose.

The estimate. The rate of good drafts is estimated as a Beta posterior:

p(good rate) = Beta(1 + good, 1 + bad)

The prior is flat, so only the person's own labels count. Counts decay with a 90-day half-life.

Certification. A kind is offered autopilot when both conditions hold:

labels ≥ 20
10% quantile of Beta(1 + good, 1 + bad) ≥ 0.90

In words: the engine is 90% confident that at most 10% of these drafts would need your edits.

• With no bad labels, the condition is met after 21 unchanged sends in a row: 0.1^(1/22) = 0.9006. About 22 are needed in practice, because older labels decay.
• The offer is one click in the queue. Nothing graduates by itself.
• Graduating sets that kind to Autopilot, switches autopilot on, and resets the demotion state.
• It is refused in regulated mode.

Spot checks. After graduation, some would-be automatic sends go to the person instead, so that labels keep arriving:

• 1 in 5 before a kind has 60 labels;
• 1 in 20 after that.

Demotion. Any one of these puts a kind on probation, and autopilot holds it:

• An anytime-valid e-process. It tests "bad rate ≤ 10%" against "bad rate = 20%". The update is:

  e ← e × 2      on a bad label
  e ← e × 8/9    on a good label
  

  A kind is demoted when e ≥ 20. By Ville's inequality, a kind whose true bad rate is at most 10% is wrongly demoted with probability at most 5%, however long it runs. For example:

  • 10 bad labels in 40 gives e ≈ 29.9, which demotes;
  • 9 in 40 gives e ≈ 13.3, which does not.

• A cancel streak: three automatic sends stopped in a row.

• Mostly rewritten: after at least 3 labels, the posterior mean of the good rate is below 0.5.

Loop 2: lessons from your edits

• When a sent draft was edited by more than 5%, the model compares the draft with what was sent. It states the durable preference in plain language, for example "Propose a specific day and time instead of asking for availability."
• The model is shown the lessons already known for that kind of email, with their ids. When an edit shows a preference a known lesson already states, even in different words, it returns that lesson's id and the lesson is reinforced rather than duplicated. Shared wording (Jaccard ≥ 0.5 over content words) is a fallback, and one edit can reinforce a given lesson only once.
• A lesson becomes active once it has been seen twice, or when you confirm it. You can retire any lesson.
• Active lessons go into every draft prompt in their context. The prompt also includes the three sent emails most similar to the new one, as examples of tone.
• Lessons shape wording only. They never authorise a fact or relax a rule.

Loop 3: outreach experiments

Arms. There are two dimensions:

Dimension	Arms
Opening angle	Specific observation · relevant question · outcome first · very short · why now
Send time	Morning (7–11) · midday (11–14) · afternoon (14–19), local time, within your sending hours

"Why now" is a sleeping arm. It is offered only for a lead whose Form D was filed in the last 90 days, and it never overstates the raise: a Form D reports the amount sold to date, not necessarily a new round.

Outcomes.

Outcome	Counts as
A human reply within 14 days	A reward
An opt-out	A failure, and is also shown as an opt-out
An automated reply	Nothing (ignored)
No reply after 14 days	A failure

Each email is settled exactly once.

Delayed feedback. A send that is still waiting counts as a fraction of a failure:

F(t) = 1 − exp(−t / 36 hours)

This assumes half of replies arrive within a day and 86% within three. The engine learns from day one instead of waiting out the reply window.

Prior. Each arm's prior is empirical Bayes, built from other accounts' settled outcomes on that arm (leaving this account out):

• It is used once those accounts have at least 30 sends on the arm.
• Otherwise the prior is centred on the pooled rate across all arms.
• An arm nobody has tried therefore starts near real reply rates, not at a flat 50%.
• The prior is Beta(1 + 20r, 1 + 20(1 − r)).

Choice.

• Arms are chosen by Thompson sampling, with 10% of decisions made uniformly at random.
• Each send logs its propensity, 0.1 / K + 0.9 × P(best), so strategies can be evaluated offline without bias.
• Counts decay with a 90-day half-life.

What the person sees

Agent & autopilot → "What the agent has learned" shows:

• Trust meters for each kind of email: the estimate with its 90% interval, and either the certified bound or the reason for probation.
• Lessons, with Confirm and Retire buttons.
• Arm tables with these columns:
  • sends settled, replies and opt-outs;
  • the estimated reply rate with a 90% interval;
  • the chance each arm is best (shown as "–" for a sleeping arm that has never been used).
• An error budget: automatic sends in the last 7 days, times the worst certified bad rate.

Graduation offers appear at the top of the queue.

Research basis

Method	Used for	Source
Thompson sampling	Outreach arms	W. R. Thompson, Biometrika (1933); O. Chapelle and L. Li, NeurIPS (2011); D. Russo et al., "A Tutorial on Thompson Sampling" (2018)
Delayed feedback	Pending sends as partial failures	C. Vernade, O. Cappé and V. Perchet, UAI (2017), arXiv:1706.09186
Certified error rates	Graduation	A. Angelopoulos et al., "Learn then Test", Annals of Applied Statistics (2025), arXiv:2110.01052; Trust-or-Escalate, ICLR (2025)
Strong acceptance as the label	Unchanged and no critical field changed	Ansible Lightspeed (2024)
E-processes and Ville's inequality	Anytime-valid demotion	Ramdas et al., survey of e-values, arXiv:2210.01948
Discounting	90-day half-life	V. Raj and S. Kalyani, discounted Thompson sampling (2017)
Sleeping arms	"Why now" only with a fresh Form D	Duolingo, KDD (2020)
Learning from edits	Lessons and own-writing examples	G. Gao et al., PRELUDE/CIPHER, NeurIPS (2024), arXiv:2404.15269; PROSE, ICML (2025), arXiv:2505.23815
Prompt-injection risk in mail agents	The security veto	The Trojan Hippo and EchoLeak attacks (2025–2026)

Calendar

Every calendar a person uses, read and written, and tied to Relationships (/app/calendar, full detail in docs/calendar.md).

• Providers.
  • Google Calendar and Microsoft 365 / Outlook.com (OAuth).
  • iCloud and any CalDAV server: Fastmail, Nextcloud, Yahoo, Zoho (app-specific password).
  • Any ICS link, read only.
  • Behind one provider interface, src/lib/calendar/providers/.
• The view. Day, week and agenda views of every connected calendar, merged and colour-coded. On a phone it opens on the agenda.
• Its own iCalendar engine.
  • A parser and writer that keeps files intact on a round trip.
  • Recurrence expanded on the event's own clock (RRULE, EXDATE, RDATE, RECURRENCE-ID overrides), so a weekly 09:00 stays at 09:00 across daylight saving.
  • IANA, Windows and custom VTIMEZONE zones.
• Linking.
  • Attendees link to contacts by address, and meetings link to deals by company domain or by the deal's name in the title.
  • Contacts and deal cards show "next meeting" and "last met", with a Schedule button.
• Meeting briefs.
  • Any meeting opens a free brief from YouBank's own data: recent emails, deal stage, the company in the directory and the terminal, and recent news.
  • The AI brief is written on a click (premium).
  • An opt-in morning run briefs the day's external meetings (premium).
• Scheduling.
  • Free times come from live free/busy across all calendars, inside the person's working hours.
  • Invitations are sent by their own provider. A video link can be pasted, or Google Meet or Teams can add one.
  • Meetings can be moved or cancelled, one occurrence or the whole series.
  • With smart scheduling (premium), the email agent offers real free times in drafts about meeting.
• Sync.
  • Incremental: Google sync tokens, Graph delta, CalDAV sync-collection, ICS ETags.
  • Google and Microsoft push changes within seconds, and polling runs as the fallback.
  • The sync is free work with no AI.
• For the meeting copilot. getMeetingAt, getUpcomingMeetings and GET /api/calendar/upcoming are a stable API, documented in docs/calendar.md.

Meetings: the meeting copilot

The copilot listens to meetings and files what it learns into Relationships (the full design, and where a calendar plugs in, are in docs/meetings.md). Two ways to listen:

• The desktop app (free). When a Zoom, Teams, Google Meet, Webex or Slack huddle call starts, it offers to start; nothing records until the person agrees to a prompt reminding them that some places require everyone's consent (with a notice to paste in the chat). It records the microphone and what the computer plays (WASAPI loopback, CoreAudio's tap on macOS 14.2+, the PulseAudio/PipeWire monitor on Linux; microphone only where that is not available), shows a red dot in the tray and a small pill with the time and Stop, and sends 30-second chunks to be transcribed. Audio is deleted once transcribed.
• The notetaker bot (Deal Team, meetings.bot, about $0.65 an hour through Recall.ai): sent to a meeting link by a click, or by an auto-join rule the person switched on; it joins as a participant and its transcript names each speaker.

Transcription uses OpenAI's speaker-labelled model when OPENAI_API_KEY is set, otherwise the ML service's Whisper and Parakeet. During a call, the copilot window shows the rolling transcript and what YouBank knows about the people in it (contacts, deal stage, recent emails, a listed company's numbers), free; live suggested questions are Pro (meetings.live), on only while switched on for that meeting.

When the meeting ends, the notes (free for three meetings a month, then Pro, meetings.notes): a summary, decisions, action items with owners and dates, open questions, each person's sentiment, interest and signals, and follow-up drafts. They are filed the way the email agent files mail: each participant matched to a contact (or added and marked for review), a meeting entry on each contact's and deal's timeline carrying their topics (contact knowledge), stated deal terms as proposed changes to accept or reject (never written over), follow-ups in the review queue (never sent by autopilot). Relationships → Meetings lists the meetings, and each meeting's page shows the transcript, the notes, the proposed updates as cards, and an "ask about this meeting" box.

Teams and live collaboration

• Teams (/app/team) are shared workspaces.
  • There are four roles: owner, admin, member and viewer.
  • People join by invitation.
  • A team always keeps at least one owner. A sole owner can make someone else an owner and then leave.
• Live collaboration (/app/collab) lets two people open a tool and work on the same model.
  • Changes appear for the other person. Someone who joins later picks up the model where it stands.
  • It runs on Postgres and Server-Sent Events, because Vercel cannot hold a socket open.
  • An append-only event log's serial id doubles as the sequence number. Streams rotate every 45 seconds and resume with Last-Event-ID.
  • Patches merge on the server with state || patch::jsonb, so the last write wins per field. Keystrokes are coalesced into one patch per 250 ms.
  • Presence is a heartbeat with a 25-second window.
  • A session shared with a team is open to that team; an unshared one stays with whoever started it.

YouBank for desktop

A desktop app for Windows, macOS and Linux (Tauri v2, in desktop/; its own README covers building, releasing and signing). Download it from /download, which picks the installer for the visitor's system from the latest desktop-v* GitHub Release and says in plain words what the app may touch.

• The site in its own window, with a tray menu, native notifications, youbank:// links (a ticker, a deal, a Studio document, a quick ask question), one instance at a time, a remembered window, an offline screen that retries, and updates from GitHub Releases.
• Quick ask from any app on a global hotkey (Ctrl+Shift+Space, ⌘⇧Space on a Mac): the terminal's assistant in a small floating window, on the same daily AI allowance.
• A local agent, each part off until the person agrees to a screen saying what it does:
  • Local files. Folders of CIMs, models and memos become Edge documents. Files are fingerprinted on the computer and only new or changed ones are uploaded; the server gets a hash of each path, never the path. The first 25 files are free; more need Pro (desktop.folders).
  • Excel and PowerPoint. Pull a Studio model as a real .xlsx (or a deck as .pptx); a saved workbook goes back to Studio as one undoable change ("Synced from desktop: 14 cells"), through the same diff as the Excel add-in, and is refused if Studio moved on meanwhile. AI edits to a local file (Pro, desktop.office_agent) run the Studio agent and write the result back, keeping a backup.
  • Alerts. Edge findings and other bell alerts, the email agent's questions and news about pipeline deals, polled every few minutes (read only).
  • The meeting copilot (see Meetings): notices calls, asks for consent, records the microphone and system audio, shows a recording pill and the copilot window, and says when the notes are ready.
  • Scheduled tasks from the tray: the morning brief and the email agent's status (free), the Edge brief and watch checks (Pro, desktop.background_ai). An AI task runs on its schedule only if the person switched it on for that computer; the server keeps its own copy of that switch and refuses otherwise.
• Signing in works like the Office add-in: the app shows a code, the person approves it at /desktop/connect, and the app keeps a device token (ybd_…) in the system keychain. It works only on /api/desktop/**. Connected computers are listed, and can be disconnected, in Settings → Desktop app.
• What the site may ask of the app is fixed at three commands (its version, open quick ask, open the agent's settings). The site cannot reach files, programs or the keychain through it.

Edge: precedents, alternative data and sale odds

Edge's next round (docs/edge-next.md) adds seven features on shared foundations (an entity crosswalk, a weekly signal store with change points that state their false-alarm rate, and a forecast ledger whose probabilities are scored when their windows close, shown under Track record). All run on free public data; the paid layers follow docs/premium.md.

• Deals (the Precedent Engine). A precedent-transactions database built from EDGAR's own merger documents: merger proxies, information statements, 14D-9s, S-4s and 8-K merger agreements with their press releases, found daily and backfilled from 2015. The small model reads one section at a time (background, opinion, termination fees, projections, approvals); every term is stored with its verbatim quote, re-read from that quote by number parsers, checked against the target's as-filed XBRL figures and prices, and given a confidence per field family (a stated prior until a family has 60 hand-labelled fields). "Deals like this one" blends meaning (embeddings of the target's business and rationale) with size, industry, buyer type, year and consideration, says why each matched, and gives the EV/EBITDA range with a split-conformal 80% band whose coverage is tested on held-out recent deals. The deal page shows the sale process as a timeline, each term with its quote and checks, and the fairness ranges as a football field. Reading the backfill is platform spend and runs only while an administrator has it on (EDGE_BACKFILL_DAILY_USD, or the switch in Deals); anyone can read one deal on demand from their own allowance. PREC in the Terminal shows the engine's set with live EDGAR search as "Find more"; the deals.precedents canvas block feeds memos and Studio.
• Pulse (the Alt-Data Pulse). Per company, from the crosswalk's ids: job postings by department and tell from Greenhouse, Lever and Ashby boards (counts and titles only), weekly Wikipedia views, federal contract awards from USAspending, WARN layoff notices (Texas, California, New York), and patents from PatentsView. Changes are flagged at a change probability of 0.9 with a large move, and each finding states that probability, the detector's measured false-alarm rate and its coverage ("1 of 3 known job boards"). Tells (integration hires, corporate development, restructuring, IPO readiness at private companies, new metros, plant build-outs) become feed cards and CRM signals. PULSE in the Terminal, a strip on /app/vc startups, and the pulse.signals canvas block.
• Deal Radar. A model probability that a company announces an agreement to be acquired within 12 months, always beside its sector's base rate and its Venn-Abers interval, with its reasons (each linked to its filing: a 13D, an officer change, a rights plan) and what would move it. A discrete-time hazard model on public filings only (XBRL as filed, holder and 8-K filings, the industry's deal wave), LightGBM on the ML service or a logistic model without it, validated out of time against the base rate and a Palepu-style logit. This round scores energy, technology and healthcare companies (EDGE_ODDS_UNIVERSE=all widens it to every US-listed operating company). Forecasts for watched companies and the top 200 go into the ledger and are scored automatically. in the Terminal and the canvas block. It is a statistical estimate from public data, not a rumour or advice, and says so wherever it appears.

Quant: the quant desk in your pocket

Quant (docs/quant-edge.md) puts the methods a quant desk uses on a person's own portfolio and ideas, with an interval and a source on every number. It lives at /app/quant ("Quant" in the navigation, after Fund copies), in the Terminal, in the AI assistant, on the Edge canvas and in the Edge feed. It produces signals, target portfolios, risk budgets, cost estimates and integrity grades for the algo engine; it never places an order. Nothing is phrased as advice: explanations come from templates in src/lib/quant/words.ts, a test checks every template and AI tool description against a list of advice words, and results carry "hypothetical" and "not investment advice" disclosures. The terms are accepted once (quant_consents).

• Backtest check (BTX). Grades any backtest A to F: the deflated Sharpe ratio with the effective number of variants tried (every backtest run is logged as a trial in quant_trials, append-only), the probability of backtest overfitting by CSCV, minimum track record length, the Harvey-Liu haircut, cost sensitivity and the breakeven cost, year-by-year stability, a block placebo, a lone peak among the parameters, and look-ahead, survivorship and language-model hygiene. "Check a claim" works from typed numbers alone. Copy trading's clone backtests are logged as trials of the fund's family and carry the grade. A D or F needs a typed, logged override before the algo engine trades it live.
• Factor Lab (FACT). A portfolio's betas to market, size, value, momentum, quality and low volatility with 95% intervals over one and three years, alpha with its t-statistic and how likely luck is, the factor share of risk, the closest mix of index funds, and factor crowding (valuation spread, comovement, 13F ownership). Factors are YouBank's own, built point in time from SEC filings and daily prices (Fama-French 2x3 sorts on the top-1,000 universe); until they have enough history, ETF proxies stand in, and the screen says which. Kenneth French's data is never shown or redistributed.
• Portfolio risk (PORT, with new tabs). Risk by holding and factor (Ledoit-Wolf and a factor model), one- and ten-day value at risk and expected shortfall by filtered historical simulation with Kupiec and Christoffersen backtests, replays of 2000, the 2007 quant crunch, 2008, 2020 and 2022 (holdings without prices then replayed through their factors), shocks such as "market -20" with the other factors drawn conditionally, the gentlest scenario that loses a given amount, drawdown odds and days to exit.
• Sizing (KELLY). Kelly on the person's own numbers as a range, the growth curve, the chance of halving at full, half and quarter Kelly, volatility targeting, risk parity, and pod-style drawdown rules simulated on a history (with the times they de-risked near a bottom). A saved risk budget is what the algo engine reads; it may adopt or tighten it, never loosen it.
• Trading costs (TCA). A pre-trade estimate for any symbol and size (Corwin-Schultz and Abdi-Ranaldo spreads, square-root impact on 20-day volume), the person's own fills from a broker CSV (Schwab, Fidelity, Robinhood, Webull, IBKR, E*TRADE) or, with Pro, a connected broker, measured against each day's bar and projected to a year, and brokers' Rule 605 and 606 reports with the payment for order flow on the person's own orders. Fills are never pooled across people and can be deleted.
• Insiders (INSC). Form 4 and Schedule 13D every ten minutes of EDGAR's day: opportunistic insider clusters (routine traders, plan trades, offerings and 10%-owner funds left out), large CEO and CFO purchases, and new or raised activist stakes, each with the base rate of what followed similar events. Cards in the Edge feed for watched companies (, ).

Underneath: a point-in-time store (an NYSE calendar with special closures; daily bars in R2; SEC company facts known from their acceptance time; a top-1,000 US universe rebuilt weekly, flagged survivorship-biased because delisted history needs a vendor), a statistics kit with known-answer tests against statsmodels, NumPy and the papers' worked examples, and an event-study engine. The nightly pass (Inngest quant-nightly, or /api/cron/quant without Inngest) refreshes it in slices. The algo engine's entry point is src/lib/quant/api.ts; the contract is src/lib/quant/contracts.ts. AI tools: quant_factor_exposures, quant_risk, quant_backtest_check, quant_signals, quant_event_base_rates, quant_costs, quant_kelly; canvas blocks: Factor Lab, Portfolio risk, Backtest check, Event base rates, Insider clusters. Free covers the cores; Pro adds depth (200 holdings, ten years, broker holdings, full integrity reports, risk budgets, broker sync, the full insider board and Item 4 readings); Deal Team adds team drill leaderboards.

Premium features

Premium features are built and switched on per plan: the plan decides who may use one, and for the paid data and models a key in the environment decides whether it can run at all. They spend money only when someone whose plan includes them uses them on purpose (a click, a ticked option); crons, monitors, prefetches and page loads always get the free method. Administrators (ADMIN_EMAILS) may use everything. Locked features stay visible with their plan badge, and trying one shows the plan's message in line with a link to the plans. The contract is in docs/premium.md; the registry is src/lib/billing/features/ (one file per area), with each feature's cost per use for the pricing model. Why each of Edge's newest features is free or premium, and on which plan, is in docs/pricing.md.

Feature	Plan	Where	What it adds
Premium reranking	Pro	Edge · Documents	Voyage rerank-3 (or Cohere Rerank 4) on every question, instead of the free reranker
Stronger answer model	Pro	Edge · Ask, "Stronger model"	EDGE_ANSWER_MODEL (e.g. gpt-5.6-sol) writes that answer
Exact-span citations	Pro	Edge · Ask, "Exact-span citations"	Anthropic Citations: quotes cut from the passages by the API
Ask across companies	Pro	Edge · Ask, "Each company separately"	One question answered from each of up to six companies' filings, side by side
Speaker-labelled transcripts	Pro	Edge · Library, on upload or per recording	OpenAI gpt-4o-transcribe-diarize; MP3 and WAV of any length, other formats to 24 MB
Hard PDFs read by LlamaParse	Pro	Edge · Library, on upload or per document	LlamaParse v2 (agentic tier) with tables kept whole
Sharper satellite imagery	Pro	Edge · a plant's panel	Planet's PlanetScope and SkySat scenes
TimesFM forecasts	Pro	Edge · Scenarios, Driver forecasts	Google's TimesFM through BigQuery, beside the free drift forecast
Flare volumes (Nightfire)	Enterprise	Edge · a plant's panel	VIIRS Nightfire flares with an estimated volume (licensed data)
GPU retraining of the deal model	Enterprise	Edge · Networks	One retraining on a Modal T4, wider and longer (the weekly one stays on CPU)
Re-read a deal with the stronger model	Pro	Edge · Deals, a deal's page	EDGE_REREAD_MODEL (e.g. gpt-5.6-sol) re-reads every term, re-quoted and re-checked; the earlier reading stays in the audit trail
Precedent exports	Pro	Edge · Deals	CSV of more than 50 precedents, with each term's quote and source (50 are free)
Second-provider claim check	Pro	Edge · Ask, "Cross-check"	The other provider's model judges each claim against its passages
Daily Pulse

Settings → Labs → Edge premium upgrades lists each upgrade with whether YouBank has it switched on and whether your plan includes it; administrators also see its cost to us, the settings behind it (names only) and the steps to switch it on.

Site, onboarding and styles

• The landing page (/) includes:

  • an auto-playing terminal demo;
  • a mock of the relationships queue;
  • a live adaptive-engine demo running the product's own code: an earned-autonomy simulator, a Thompson sampling simulator and lessons;
  • live demos: trading comps, an assistant answer with citations, and the startup directory;
  • segments, the planned plans, a style gallery, data and method, and an FAQ.

• Per-career pages at /for/<role>.

• Sign-in with Google through Better Auth (self-hosted, its tables in our database), then a six-step onboarding survey: role, group, experience, focus, style and goals.

• 14 styles:

  • Dark: Terminal, Midnight, Aurora, Graphite, Phosphor, Evergreen, Nordic.
  • Light: Daylight, Paper, Ivory & Gold, Ocean, Rosé, Solar, Lavender.

  Components use design tokens only. scripts/gen-themes.ts regenerates src/app/themes.css from src/lib/themes.ts.

• The top bar and the sidebar. Each person pins the features they use to the top bar, in their own order, with names or icons only (icons only on a tablet, so the bar fits); the page they are on shows as a dashed tab when it is not pinned. The sidebar (the button at the left, or ⌘/Ctrl+\) holds every feature, the workflows inside them (terminal functions, Newsroom views, Relationships tabs, settings) and the tools picked for their role, with search across all of it. It opens over the page or docks beside it. When a workflow's page is already open, it switches view in place, so open panels stay put. Saved in profiles.extra.nav (src/lib/nav.ts).

• Built-in controls. Dropdowns (src/components/ui/Select.tsx) and confirmation and text-entry dialogs (src/components/ui/Dialog.tsx) replace the browser's own: themed like everything else, animated, with keyboard support, search in long lists, and type-to-confirm for deletions that affect a whole team.

• On a phone, everything works. Nothing is desktop-only; each screen is laid out again to fit.

  • Navigation. The first four pinned features become a tab bar along the bottom edge, and More opens the sidebar as a bottom sheet: the same search, features and workflows, at finger size, plus an arrange view that shows which pins are tabs. The top bar slims to the page's name, alerts, style and account. Pinning and arranging are one list for both, saved the same way.

---

Architecture

     Browser: Next.js App Router pages, React 19      Excel and PowerPoint: the add-in (Office.js)
                                        │                     │  device token, same /api/studio routes
                                        ▼                     ▼
  Vercel ─── Next.js 16 server: pages and route handlers under /api/*
   │   Cron 06:00 UTC ─► /api/cron/sync     startup directory and Form D refresh
   │   Cron 11:00 UTC ─► /api/cron/agent    morning agent run (drafts only)
   │
   ├──► Neon Postgres (Drizzle ORM, postgres.js through the pooler; also runs on Supabase)
   ├──► OpenAI Responses API  /  Anthropic Messages API
   ├──► SEC EDGAR (submissions, XBRL company facts, full-text search) · Financial Modeling Prep
   └──► Mailboxes: IMAP + SMTP (app password)  or  Gmail API (OAuth)

  Neon Function "autopilot" ◄── schedule trigger (every 20 min)
        └──► POST /api/cron/autopilot  (Authorization: Bearer AUTOPILOT_SECRET)
  Neon Function "news" ◄── schedule trigger (every 20 min, a minute later)
        └──► POST /api/cron/news       (the Newsroom: sources, stories, alerts, briefs)

The mail loop:

new mail ─► sync (Gmail history, or IMAP) ─► resolve thread ─► triage ─► draft ─► consider
   ├─ held      ─► review queue, with the reasons
   └─ scheduled ─► hold + sending hours ─► re-read live thread ─► sendDraft ─► learnFromSend (loops 1–2)

reply arrives ─► recordReplies ─► stop the sequence / record an opt-out ─► settleOutreach (loop 3)

src/lib/crm/, file by file:

File	Role
model.ts	Vocabulary: modes, stages, categories, contact and draft kinds, consent domains
agent.ts	Triage and drafting prompts and schemas
sync.ts · mailbox.ts · imap.ts · gmail.ts	Mailbox access, incremental changes, thread resolution, sending over SMTP or the Gmail API
accounts.ts · crypto.ts	Connected mailboxes; AES-256-GCM encryption of credentials
db.ts	Thread ingestion, triage processing, draft creation, reply recording, activity
send.ts	sendDraft(), the only code that sends
autopilot-rules.ts	Pure rules: settings, autonomy levels, sending window, send verdict, automated-mail detection
autopilot.ts	considerDraft, the heartbeat tick, the send queue, answering questions
engine-rules.ts · engine.ts · learning-math.ts	The adaptive engine
knowledge.ts · settings.ts	Playbook and questions; settings, voice, per-person lock
scan.ts · actions.ts	Pure scan rules (follow-ups, stale deals, nurture, funding signals); suggested actions
nurture.ts · campaigns.ts · outreach.ts · contacts.ts · compose.ts	Nurture rules, campaigns and leads, contacts, new emails
run.ts	The morning run

---

Data model

All tables are in src/db/schema.ts.

Area	Tables
Accounts and research	profiles, workflow_runs, documents, peer_groups, peer_group_members, comps_sheets, manual_inputs, footnotes, company_cache, kv_cache, startups
Teams	teams, team_members, team_invites
Live collaboration	collab_sessions, collab_events, collab_presence
Relationships	email_accounts, crm_contacts, crm_deals, crm_threads, crm_messages, crm_drafts, crm_questions, crm_playbook, crm_settings, crm_actions, crm_signals, crm_nurture_rules, crm_nurture_log, crm_campaigns, crm_campaign_leads
Adaptive engine	crm_trust (per stratum: good, bad, observations, unchanged, e-process, cancel streak), crm_arms (per arm: decayed pulls, rewards, negatives), crm_lessons, crm_learning_events (every label and outcome, for audit and offline evaluation)
Studio	studio_docs (workbook, deck and comments as JSON, each edit an atomic jsonb update), studio_events (every patch with its undo; the serial id is the live-stream cursor and the add-in's sync cursor), studio_runs (each agent run: instruction, status, summary, stats), studio_checkpoints (named snapshots of the workbook and deck)
Excel and PowerPoint	office_pairings (a code awaiting approval: the hashed poll secret and the expiry), office_devices (connected installs: the hashed token, last use, revocation)
Meetings	meetings (each meeting: source, consent, participants, the live switch, notes), meeting_chunks (the transcript, one row per audio chunk), meeting_links (contacts and deals), meeting_settings; a meeting's proposed changes are rows with

Migrations.

• On a fresh database, drizzle-kit push creates every table from the schema.
• Changes to an existing database are hand-written, idempotent SQL in drizzle/, applied in order with scripts/apply-sql.mts:
  • 0001_teams
  • 0002_crm
  • 0003_collab
  • 0004_outreach
  • 0005_autopilot
  • 0006_adaptive
  • 0007_studio
  • 0008_office
  • 0009_ai_usage
  • 0010_contact_knowledge
  • 0016_crypto
  • 0017_desktop
  • 0018_calendar
  • 0019_meetings
  • 0023_copytrade
  • 0024_seat_cooldown
  • 0026_quant
• After applying them, drizzle-kit push --strict should propose nothing (answer "No" to anything it proposes). Never run drizzle-kit push without --strict (or with --force) against an existing database: indexes the SQL files create are not all in schema.ts, and push would drop them.

---

Tech stack

Layer	Choice
Framework	Next.js 16.3 (App Router, Turbopack), React 19.2, TypeScript 5
Styling	Tailwind CSS v4 with design tokens, 14 generated themes, motion, lucide-react icons
Database	Neon Postgres (free plan) through postgres (postgres.js) and the pooler; portable to Supabase (scripts/db-setup.mts)
ORM	Drizzle ORM 0.45, drizzle-kit
Auth	Better Auth 1.6 (self-hosted): email and password, plus Google when configured; no emailed reset yet (an admin resets); tables auth_users, auth_sessions, auth_accounts, auth_verifications
AI	OpenAI Responses API (openai 7), Anthropic Messages API (@anthropic-ai/sdk), zod 4 for structured output
Mail	imapflow (IMAP), nodemailer (SMTP), mailparser (MIME), Gmail REST API
Data	SEC EDGAR (submissions, XBRL company facts, full-text search), Financial Modeling Prep, Y Combinator, a16z, Thiel Fellows, Show HN
Hosting	Vercel (functions and Cron)
Scheduling	Vercel Cron (daily jobs) and a Neon Functions schedule trigger (the five-minute heartbeat)
Statistics	In-house, dependency-free: Beta CDF and quantile, Gamma and Beta sampling, Thompson sampling
Tests	tsx scripts: unit suites, plus end-to-end runs against a Neon branch and an Ethereal mailbox

---

Running locally

pnpm install
cp .env.example .env.local                 # fill in the keys (see below)
set -a; . ./.env.local; set +a
pnpm exec drizzle-kit push                 # create the tables on a fresh database
pnpm dev                                   # http://localhost:3000

Development tips:

• Use a Neon branch rather than production for development.
• Outside production, YOUBANK_DEV_USER=<name> signs you in as dev-<name> without Google. It is useful for screenshots and tests against a branch.

Other commands:

pnpm exec tsx scripts/gen-themes.ts                               # after editing themes
pnpm exec tsx --env-file=.env.local scripts/snapshot.ts           # refresh the landing page's demo data
pnpm exec tsx --env-file=.env.local scripts/run-workflow.ts <id>  # one live workflow run
pnpm exec tsx --env-file=.env.local scripts/backfill.ts           # startup directory backfill
bash scripts/preflight.sh                                         # everything that must pass before a deploy

---

Environment variables

Variable	Required	Purpose
DATABASE_URL	yes	Pooled connection string (Neon's -pooler host, or Supabase's transaction pooler); no channel_binding parameter
DATABASE_URL_UNPOOLED	for schema pushes	Direct connection (Neon) or Supabase's session pooler; drizzle-kit push and scripts/db-setup.mts need it
BETTER_AUTH_SECRET	yes	Signs session cookies and encrypts stored OAuth tokens (openssl rand -base64 48)
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, NEXT_PUBLIC_GOOGLE_SIGNIN=1	no	Google sign-in alongside email and password: an OAuth client (Web) with redirect URI <site>/api/auth/callback/google; all three or none
BETTER_AUTH_URL	no	Sign-in's base address when it is not NEXT_PUBLIC_SITE_URL (local development)
AUTH_SESSION_DATA_TTL	no	Seconds the session cookie is trusted before a database read (default 900)
AI_PROVIDER	yes	openai or anthropic
OPENAI_API_KEY or ANTHROPIC_API_KEY	yes	The key for the chosen provider
OPENAI_MODEL, OPENAI_RESEARCH_MODEL, OPENAI_REASONING_EFFORT, ANTHROPIC_MODEL	no	Defaults: gpt-6-astra, gpt-5.4-mini, medium, claude-opus-5
FMP_API_KEY	yes	Prices and market caps
MARKET_BACKUP	no	off turns off the market-data backups (Nasdaq, the ECB rates, CoinGecko, AI research); on by default
MARKET_RESEARCH_MODEL, MARKET_RESEARCH_ESCALATION_MODEL	no	The models AI research uses; defaults gpt-5.6-luna, retrying with gpt-5.4-mini

---

Connecting a mailbox

With an app password (any of the presets)

1. Create an app password with your provider:
   • Google: https://myaccount.google.com/apppasswords. The account needs 2-Step Verification.
   • iCloud: an app-specific password under Sign-In and Security.
   • Yahoo, Zoho and Fastmail: their equivalents.
2. In YouBank, open Relationships, choose the provider, and paste your address and the app password.
3. YouBank verifies the IMAP and SMTP logins before saving anything, then encrypts the password. Custom servers must be public addresses on standard mail ports (IMAP 993 or 143; SMTP 465, 587, 25 or 2525).

With the Gmail API (OAuth)

1. In Google Cloud, enable the Gmail API, then create an OAuth client ID of type Web application.
2. Add https://<your-domain>/api/crm/gmail/callback as an authorised redirect URI. Add the localhost equivalent too, for local work.
3. Set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET and EMAIL_TOKEN_SECRET.
4. Open Relationships and press Connect Gmail.

How the OAuth connection behaves:

• The agent asks for gmail.readonly and gmail.send only. It cannot modify or delete anything in the mailbox.
• The consent flow carries a nonce in an httpOnly cookie.
• Refresh tokens are encrypted before storage. A refresh failure marks the mailbox needs_reauth and gives the reason.
• Google classes both scopes as restricted. Your own account and a test list of up to 100 users work immediately. Serving anyone else requires Google's verification review.

---

Connecting a calendar

Settings, Calendar. The steps for each provider's developer console are in docs/calendar.md. In short:

• Google: enable the Calendar API on an OAuth client (the Gmail one works). Add https://<site>/api/calendar/oauth/google/callback as a redirect URI. Set GOOGLE_CALENDAR_CLIENT_ID and _SECRET, or rely on the Gmail client.
• Microsoft: register an app for any organisation and personal accounts. Add https://<site>/api/calendar/oauth/microsoft/callback. Grant Calendars.ReadWrite, User.Read and offline_access. Set MICROSOFT_CLIENT_ID and MICROSOFT_CLIENT_SECRET.
• iCloud, Fastmail, Nextcloud, Yahoo, Zoho: nothing on the server. People pick the service and paste an app-specific password. YouBank signs in and lists the calendars before saving anything.
• ICS links: paste any webcal:// or .ics link. It is read only.

Every provider needs EMAIL_TOKEN_SECRET. Without it each shows "Not set up yet".

Deploying

The production runbook is docs/production.md: which variables are required in which Vercel environment, the first deploy, the custom-domain checklist (OAuth redirects, the Stripe webhook, BETTER_AUTH_URL, Inngest), health checks, what to do when a dependency is down, and the Hobby plan's limits.

The Vercel project is connected to this repository: every push to main (a merged pull request) deploys to production, and pull requests get preview deployments. Apply new migrations before or right after the deploy that needs them (pnpm exec tsx scripts/db-setup.mts finish replays every drizzle/*.sql).

1. Environment. Set the variables above in the Vercel project, not in .env.local.
2. Database. Apply any new migration to production, then check for drift:

   set -a; . ./.env.local; set +a
   DATABASE_URL="$DATABASE_URL_UNPOOLED" pnpm exec tsx scripts/apply-sql.mts drizzle/0011_newsroom.sql
   pnpm exec drizzle-kit push --strict   # a drift check: it should propose nothing; answer "No" to any prompt
   

3. Build and ship:

   bash scripts/preflight.sh         # themes, typecheck, lint, tool packs, production build
   vercel --prod --yes
   

4. Crons. vercel.json schedules /api/cron/sync (06:00 UTC), /api/cron/agent (11:00 UTC), /api/cron/calendar (06:15 UTC) and /api/cron/calendar-briefs (10:45 UTC).
5. The autopilot heartbeat. Deploy it on the production branch with the Neon CLI. It is deployed this way, not from neon.ts, because it carries secrets:

   neon functions deploy autopilot --src neon/autopilot.ts \
     --env YOUBANK_URL=https://<your-domain> --env AUTOPILOT_SECRET=<secret>
   neon triggers create --function-slug autopilot --name autopilot-heartbeat --cron '*/5 * * * *'
   

6. The Newsroom heartbeat, the same way:

   neon functions deploy news --src neon/news.ts \
     --env YOUBANK_URL=https://<your-domain> --env AUTOPILOT_SECRET=<secret>
   neon triggers create --function-slug news --name news-heartbeat --cron '*/10 * * * *'
   

   The calendar heartbeat runs every fifteen minutes. It calls the free sync, then the morning briefs, which run only for people who switched them on. Apply drizzle/0018_calendar.sql first:

   neon functions deploy calendar --src neon/calendar.ts \
     --env YOUBANK_URL=https://<your-domain> --env AUTOPILOT_SECRET=<secret>
   neon triggers create --function-slug calendar --name calendar-heartbeat --cron '*/15 * * * *'
   

7. Billing (Stripe). Create the products and prices, and the webhook endpoint, from PLANS:

   STRIPE_SECRET_KEY=sk_live_... pnpm exec tsx scripts/stripe-setup.ts --live \
     --webhook https://<your-domain>/api/billing/webhook
   

   • Set the printed STRIPE_* lines in Vercel.
   • In the Stripe dashboard, switch on the customer portal and allow plan changes between these prices.
   • Run it first with a test key (sk_test_..., no ) on a preview deployment.

• curl https://<your-domain>/api/health/live answers at once without touching the database (for uptime monitors).
• curl https://<your-domain>/api/health should say "status":"ok". It answers degraded (200) when the app serves but something is missing (a migration, a production setting, an optional service half set up) and down (503) without a database, its core tables, or a setting every page needs. Operators also see which migration is missing which tables. Anyone sees the status; the names of missing settings and which services are on need the cron secret (-H "Authorization: Bearer $CRON_SECRET") or an ADMIN_EMAILS session. No value is ever shown.
• curl -i https://<your-domain>/api/cron/autopilot should return 401 without the secret.
• After five minutes, the Vercel logs should show POST /api/cron/autopilot.

---

Testing

Suite	Command	What it covers
Learning (63 tests)	pnpm exec tsx scripts/test-learning.ts	Beta CDF and quantile against exact values; certification boundaries; the e-process; scoring and critical fields; decay; Thompson sampling and propensities; delayed feedback; the security veto
Autopilot (62 tests)	pnpm exec tsx scripts/test-autopilot.ts	Settings normalisation, autonomy levels, sending windows and time zones, the send verdict, placeholders, automated-mail detection, the daily cap, playbook selection
Outreach (53 tests)	pnpm exec tsx scripts/test-outreach.ts	Follow-up, stale-deal and nurture candidate rules; funding-signal matching; company normalisation; sequence steps; the consent guard
Calendar (119 tests)	pnpm exec tsx scripts/test-calendar.ts	iCalendar parsing, folding and round trips; RRULE expansion (ordinals, BYSETPOS, BYWEEKNO, UNTIL, COUNT, jumping to the window); time zones and DST, including custom VTIMEZONE, Windows and Mozilla zone names and floating times; editing one occurrence (EXDATE, overrides); attendee-to-contact and deal matching; slot finding; video links; the private-network guard; the Google, Microsoft, CalDAV and ICS adapters against fixtures in scripts/fixtures/calendar/
Gmail parsing (12 tests)	pnpm exec tsx scripts/test-gmail-parse.ts	Address splitting (including quoted commas), MIME bodies, headers
Inference (58 tests)	pnpm exec tsx scripts/test-inference.ts	The command parser; Welch beta, Kupiec, Parkinson; forecasts, seasonality and nested intervals; rating tables, the Ohlson units, left-out views; knowledge tracing and its policies; Kaplan-Meier and Poisson-binomial; relationship strength, contact knowledge, deal odds and pipeline simulation; the recession probit and Sahm rule; copula rank correlation; Monte Carlo and forecasting over a live Studio workbook
Newsroom (73 tests)	pnpm exec tsx scripts/test-news.ts	URL, title and ticker cleaning; RSS, Atom and RDF; EDGAR's latest-filings feed, 8-K items and 13D pairs; Federal Register, GDELT and radar items; sector radars (FERC, DOE and NRC milestones, Fed bank applications, ITC cases, trials, FDA approvals, recalls, places and the map); robots.txt precedence; article extraction and paywall markers; classification and importance; clustering thresholds, figures (rounded or not) and filings; company names against SEC's listings; ranking reasons and mutes; desks; preferences, quiet hours and brief times; budget tiers; premiums, implied multiples and league tables; alert decisions; the calendar across daylight saving; research acceptance; tidying the model's reading; the brief email

The end-to-end scripts write rows under a throwaway user. Point them at a Neon branch, never at production.

---

Security, privacy and compliance

Credentials.

• Mailbox passwords and refresh tokens are encrypted with AES-256-GCM before storage, and decrypted only to talk to the mail server.
• Disconnecting a mailbox deletes the row, so no credential survives it.

Sending.

• There is one sending function, sendDraft().
• A single-use claim means a draft cannot go twice.
• The live thread is re-read before every automatic send.

Prompt injection.

• The security veto holds any automatic email containing a link, address or account number the person did not provide.
• Automated messages are never answered.
• Loop guards limit autopilot to two sends per conversation per day.
• Lessons cannot authorise facts.

Excel and PowerPoint.

• The add-in signs in with a device code. Only hashes of the poll secret and the device token are stored.
• A token is collected once. A code is approved only from a signed-in browser session, never by another add-in, and the approval page warns people to approve only a code they see in their own Office.
• Every install is listed with its last use, and disconnecting one revokes its token at once.
• Uploaded printouts and data-room files are read as content: the reviewer's marks are the only requests, and printed text is never followed as an instruction.

Crypto.

• YouBank never asks for, stores or uses a private key or seed phrase, and has no custody, trading or transfer feature. Connecting a wallet shares its address only.
• The one transaction it ever asks a wallet for is a notarization: zero value, to the person's own address, on Base, shown by their wallet for them to approve. The server verifies it by reading the chain; it signs nothing.
• Paid crypto APIs are called only when an entitled person clicks, never from a page load or a cron. Errors.
• Nobody sees a raw error. A failure meant for people ("This draft was already sent", a permission or a limit) is shown as written; anything that looks internal (SQL, a URL or a key, a stack trace, an HTML error page, a provider's or a parser's own text) becomes "Something went wrong on our side (ref …)", and the full error is logged as one JSON line under that reference (and sent to Sentry when SENTRY_DSN is set). One rule decides, in src/lib/error-text.ts, shared by the server (src/lib/errors.ts) and the browser (src/lib/client/errors.ts), which also screens whatever reaches it some other way (a proxy's HTML page, an empty 500, "Failed to fetch") and falls back to a plain line by status: sign in again (401), plan needed (402), no access (403), not found (404), a limit (429), our side (5xx), offline.
• Streams (the AI chat, workflows, Studio's agent, document answers) end with a screened message, never a dropped connection; errors stored for later (a canvas step, a document's reading, a run) are screened before they are written.
• A crashed screen shows a plain message, the reference (Next's digest, also in the log line) and Try again and Home. Each workspace has its own boundary, so the navigation stays; a tab left open across a deploy is told to reload.

Scheduled endpoints.

• Cron and heartbeat endpoints require bearer secrets.
• The Neon relay accepts only real trigger deliveries.

Regulated firms.

• Regulated mode turns autopilot off and keeps every send human and in the firm's own mailbox.
• The audit CSV records every draft, who sent it and how much it was edited.
• This follows FINRA's 2026 oversight report, which flags AI agents acting "without human validation and approval", and the SEC's January 2025 off-channel recordkeeping fines ($63.1M across 12 firms).

Consent and frequency.

• Campaigns skip recipients at EU, EEA and Canadian domains unless you confirm consent or another lawful basis.
• A recipient gets at most one first cold email through YouBank per 30 days.
• Opt-outs stop the sequence, and nothing but a reply is ever sent automatically to that person again.

Your data.

• No model is trained on your data.
• What the engine learns is per-account statistics and plain-language lessons, which you can read and retire.

Known gaps, stated plainly:

• Campaign emails do not yet add a postal address, an unsubscribe line or a List-Unsubscribe header.
  • CAN-SPAM requires a working opt-out mechanism and a physical postal address in commercial email.
  • Google and Yahoo require one-click unsubscribe from bulk senders.
  • Add these before running campaigns at volume.
• Gmail's OAuth scopes are restricted. Beyond 100 test users, the OAuth path needs Google's verification. The app-password path works today.
• There is no SOC 2 report yet.
• Market data is used under a developer licence. Redistribution terms need a proper licence before paid plans launch.

---

Research behind the product

AI inference (September 2026): LLM inference techniques, financial machine learning and knowledge tracing, and Bloomberg's functions against the free data that can replace them. What was built from them, the data licences and what was left for later: docs/research/ai-inference.md.

Three earlier research reports are behind the adaptive engine, the website and the plan:

• new AI and machine-learning methods;
• the fintech and dealmaking market;
• startup and venture context.

Conclusions that shaped the product:

• Breadth is not a moat; the loop is. The defensible wedge is this sequence:

  1. a Form D, 8-K or S-1 event;
  2. the company in the startup directory;
  3. who in your inbox knows them;
  4. a drafted outreach email that you approve.

  Incumbents are converging on data access. Affinity shipped an MCP server and agents in 2026. Crunchbase, CB Insights, Tracxn, Dealroom, Specter and Harmonic all shipped MCP connectors. Anthropic's Claude for Financial Services added low-cost data partners.

• Trust has to be earned and measured. Recent work shows a model's self-graded confidence is poorly calibrated. That is why autonomy is certified on the person's own decisions, with exact bounds and anytime-valid demotion.

• Pricing started in the $30–$300 per seat per month band. The cost model in docs/pricing.md moved it: model time is most of what YouBank costs, so the prices below keep at least 70% of the monthly price (65% of the yearly one) at typical use and stay positive even for someone who uses the whole AI allowance. See Plans and billing.

---

Plans and billing

Plan	Price	AI allowance per seat	For
Free	$0	$0.75 a day, $3 a month	The terminal, filings, comps and the Newsroom, Edge's free core (precedents, Deal Radar, the Pulse, Call Desk, the LBO lab, one thesis, hold-to-talk), with a taste of the assistant
Campus	Free with a .edu address	$2 a day, $10 a month	Students: the full terminal and data, AI workflows, a recruiting pack
Pro	$59 a month, or $49 a month billed yearly	$6 a day, $25 a month	Individuals: a working AI allowance, the relationships agent and one mailbox, AI meeting briefs, and most premium features: deep research and builds, Edge's premium layers (deal re-reads, claim cross-checks, the daily Pulse, the Radar screener, the LBO refine, five theses with the weekly deep pass, live voice)
Deal Team	$149 per seat a month, or $125 billed yearly; three seats minimum	$12 a day, $60 a month	Teams: autopilot and campaigns, more mailboxes, the adaptive engine across the team, shared workspaces and team theses, the meeting notetaker, Planet imagery in 3D, Dune queries, Eva-4B on calls, fund blends, assignable seats
Enterprise	$299 per seat a month, billed yearly; five seats minimum	$30 a day, $130 a month	Firms: regulated mode and audit exports, SSO and admin controls, data residency options, bring-your-own data licences (Nightfire), GPU retraining of the deal model

• Where the numbers come from. docs/pricing.md has the unit costs with sources and dates, the light, typical and heavy personas, and the cost to serve and margin per plan.
  • src/lib/billing/costs.ts holds the same model in code.
  • scripts/test-billing.ts fails if a price, a cost or a new premium feature breaks the margin target.
• Premium features. Premium features are listed in src/lib/billing/features/, one file per area, each naming the least plan that includes it.
  • Settings, under Plan, groups them by area and shows which are unlocked.
  • The contract every premium feature follows is in docs/premium.md.
• AI credit packs. $10, $25 and $50 packs add $6, $15 and $32 of AI use, used only after the month's allowance, oldest pack first; they never expire. Bought with a one-time Checkout on any plan.
• Allowance months start on the billing anniversary for subscribers (and their seat holders) and on the 1st (UTC) otherwise.
• Team seats. The buyer of Deal Team or Enterprise holds one seat and gives the others to members of a team they own or administer (Settings, under Plan). A seat count changed in the portal reaches YouBank through the webhook, which takes back the newest seats beyond it.
  • A freed seat can go back to the same person at any time, but to someone else only 30 days after it was last given, so one seat cannot pass a fresh AI allowance from person to person. The seats API names the date and the Plan tab shows it; administrators are exempt (docs/pricing.md, drizzle/0024_seat_cooldown.sql).
• Billing is Stripe.
  • POST /api/billing/checkout starts Checkout for a plan, monthly or yearly, with a seat count. Someone who already pays is sent to the portal; an open checkout for the same purchase is reused, others are expired, and the create call carries an idempotency key, so nobody ends up with two subscriptions.
  • POST /api/billing/credits starts a one-time Checkout for a credit pack.
  • POST /api/billing/portal opens Stripe's billing portal (plan switches, seats, cancellation at period end, card, invoices); lists Stripe-hosted invoices and receipts.

---

Roadmap

• Compliance for cold email:
  • a postal address and an unsubscribe line in campaign emails;
  • a List-Unsubscribe one-click header;
  • opt-outs honoured within 48 hours.
• A YouBank connector inside Claude and ChatGPT: a remote MCP server exposing Form D × directory × XBRL.
  • It would follow the stateless 2026-07-28 MCP specification, with MCP Apps for rendered results.
  • On Vercel it would use mcp-handler, which needs its own OAuth 2.1 authorisation server.
• Mailboxes: Microsoft 365 and Outlook, and Google verification (CASA) for the Gmail API path.
• Studio and the add-in:
  • single sign-on in the add-in through Office's nested app authentication, for Microsoft 365 tenants;
  • an AppSource listing, so the add-in installs from Office's store;
  • pushing changes to the add-in over a stream, instead of a four-second poll;
  • Excel charts, conditional formats and data validation in the round trip;
  • data-room files over 4 MB, uploaded straight to storage.
• Engine:
  • fit the reply-delay curve to real data once there are 200 replies;
  • evaluate new angle strategies offline from the logged propensities;
  • suggest new angles from what replies have in common.
• Company:
  • SOC 2 Type I;
  • properly licensed market data.
• Premium: speaker labels for long M4A and MP4 recordings (needs cutting on the ML service); methane cards per plan (today Carbon Mapper is a site licence and its cards are shared findings).

---

Repository layout

YouBank/
  .github/workflows/desktop.yml   builds the desktop installers into a draft GitHub Release
  desktop/                 YouBank for desktop (Tauri v2): the Rust app in src-tauri/, its own pages in src/
  docs/                    product thinking, decisions, specs, market research
  drizzle/                 hand-written SQL migrations (0001–0008)
  neon/autopilot.ts        the five-minute heartbeat relay (Neon Function)
  neon/calendar.ts         the fifteen-minute calendar sync and morning-brief relay (Neon Function)
  scripts/                 tests, end-to-end runs, migrations, snapshot, themes, backfill, preflight
  src/app/                 routes: marketing, /for/<role>, /download, /onboarding, /app/*, /office/* (the add-in),
                           /desktop/connect, /api/*
  src/components/
    crm/                   workspace, review queue, inbox, pipeline, contacts, meetings, campaigns,
                           nurture, agent settings, engine insights, mailbox bar
    calendar/              the calendar view, event panel and brief, scheduler, Calendar settings,
                           meeting lines and Schedule buttons for contacts and deals
    studio/                Studio home, workspace, grid, deck view, slide charts, state hook
    office/                the add-in's task pane, and the connect and install pages
    marketing/             landing page, adaptive-engine demo, terminal demo, role pages, live demos
    terminal/              command bar, panels, screens (DES FA COMPS PREC CAP FIL EVT INS XBRL AI PG TOOLS,
                           and the crypto screens CRYP TOKEN DEFI STBL YLD BTCN RAISE UNLK TRSY RWA WALLET)
    crypto/                the Crypto page, wallet connection, portfolio, notarizing, the mining map, Dune
    copy/                  Fund copies: fund pages, the treemap and backtest charts, copies, order review
    workflows/             tool gallery, runner, form, output blocks
    theme/  charts/  motion/  ui/
  src/db/schema.ts         every table
  src/lib/
    crm/                   the relationships agent, autopilot and the adaptive engine (see Architecture)
    calendar/              iCalendar, RRULE and time zones, the four provider adapters, sync and push,
                           linking, briefs, free/busy and slots, the meeting copilot API
    meetings/              the meeting copilot: transcription, stitching, matching, notes, filing into
                           Relationships, the notetaker bot, meeting context (docs/meetings.md)
    studio/                formula engine, functions, number formats, templates, audit, deck, edit
                           operations, persistence, .xlsx/.pptx, the Studio agent, workbook sync,
                           checkpoints, the brand check, reading printouts and data rooms
    office/                the add-in: pairing and tokens, the Excel and PowerPoint adapters, the
                           manifest, its API client
    desktop/               the desktop app's server side: pairing and tokens, the alerts feed,
                           scheduled tasks, local files, Office sync, the release lookup
    ai/                    models, config, agent (OpenAI Responses and Anthropic), data tools, prompts
    edgar/  fmp/  vc/      SEC EDGAR and XBRL, prices, startup directory and Form D
    crypto/                crypto data (CoinGecko, DefiLlama, mempool.space, public nodes, SEC frames),
                           addresses, portfolio risk, valuation, notarization, mining sites and the map
                           layer, Newsroom signals, assistant tools
    copytrade/             13F parsing and EDGAR ingestion, CUSIP mapping, clone weights, the backtest,
                           paper copies, the rebalance, SnapTrade, one-tap approval, the cron pass
    quant/                 Quant Edge: the contract and api.ts for the algo engine, the stores, the
                           statistics kit (stats/), point-in-time data (pit/), factors, event studies,
                           integrity, risk, sizing, TCA, insiders, drills, adapters, jobs, AI tools
    workflows/             tool contract, prompt builder, registry, a pack per role
    teams/  collab/  auth/ teams and roles, live collaboration, sign-in
    themes.ts calc.ts roles.ts metrics.ts company.ts

Further docs

Doc	What
docs/production.md	Running in production: environment variables per Vercel environment, first deploy, the custom-domain checklist, health, outages, Hobby limits
docs/06-product-overview.md	Routes, terminal functions, the tool system, the AI layer, theming, commands
docs/pricing.md	What each plan costs to serve: unit costs with sources, personas, margins, and the prices and AI allowances that follow
docs/calendar.md	The calendar: provider setup, sync and push, the iCalendar engine, briefs, scheduling, and the meeting copilot API
docs/premium.md	Premium features: the plans, the feature registry, entitlements and the rules every premium feature follows
docs/meetings.md	The meeting copilot: desktop capture, the notetaker bot, live suggestions, notes and how they are filed into Relationships, the calendar's place, known gaps
docs/algo-trading.md	Algo trading: what each broker officially allows (Webull, Robinhood, SnapTrade, Alpaca, IBKR, Tradier), how algo trading is done well, AI-agent guardrails, regulation, the design, setup, known gaps and what counsel must review
docs/copy-trading.md	Fund copies: 13F ingestion and CUSIP mapping, the backtest, paper and live copies, SnapTrade, safety rails, and what counsel must review before real trading