Tasu

The smartest knowledge, in an MCP.

Folders compiled from the talks, essays and posts you trust. Each idea holds one position, dated and cited. Add a public folder to your AI in one line, or open a workspace for your own.

Open tasu.ai  ·  Read the white paper  ·  Run your own

Two ways in · What it does · Why it holds up · Compared · Inside the app · Personal folder · Assets · Workspaces · Run your own · Work on it

Two ways in

	A public folder	Your own workspace
What it is	A folder its owner made public, such as Hormozi or Y Combinator	Your folders, assets and personal folder, private
How you use it	Copy its MCP address from tasu.ai into Claude, ChatGPT or Cursor	Open a workspace with its passphrase, on the web or as an app
What you get	Answers from its positions, each claim with its author and date	Everything below: drop, ask, assets, your people, your twin
What it costs	Nothing: your own AI does the thinking	The model calls, on your OpenRouter key
Who feeds it	Its owner only	You

What it does

Drop. A link, a PDF, a Word file, a video or a transcript. Each idea is filed under a concept with its numbers, signed with author and date. A claim that clashes with what you hold waits for your call.

Ask. Ask anything, at 3 levels: the answer, a lesson from zero, or guided steps. It answers from your sources only, and the last line names who said it and when.

Use. Assets keep your docs, tables and pages. A chat reads them, and proposes changes you apply in one click.

Share. Make a folder public, and anyone adds it to their AI with its own address. It reads only; you alone feed it.

Drop compiles each source once. Ask reads what was compiled. The white paper walks through both, step by step.

Why it holds up

Zero duplicates	A source dropped twice is caught on arrival. A new idea joins the concept it belongs to.
Zero hidden contradictions	Every clash is caught at drop and waits for your call. Each concept keeps one position.
Linked ideas	Each idea links to its nearest by meaning: needs, causes, supports, contradicts. A why or how question follows the links 2 steps.
Light	It keeps the claims and drops the raw transcript. Search reads a 1 KB card per concept.
Flat cost	A question reads 69,000 characters at most, at any size. A thank-you reads none.
Your data	In your own database or as markdown files. Exported in one click, restored in one. Open source, on your own key.

Compared

Most AI tools search your files for the closest passages. Tasu reads each source once and compiles it into positions, each backed by dated evidence, and asks you when two sources clash.

	Classic AI chat	Passage search (a chat over your files)	Tasu
Answers from	Its training data	4 to 6 pieces of your files	Positions merged from up to 20 sources
Proof	None, or a link it made up	The piece it found	Author and date on every claim
When sources disagree	Blended into one answer	Both pieces, unflagged	Flagged at each drop; you rule
Tables	Pasted as text	Read as text pieces	Totals over every row, up to 40,000
In your own AI	No	Inside its own app	Any public folder, by MCP
What you get	A reply	A reply	Answers at 3 levels, one-pagers in 9 languages, clashes to rule on

Inside the app

Chat

• One list. The side panel lists everything a chat can use: your personal folder, then your categories with their folders and assets, then the rest. Names only; counts sit on each folder's page.
• Pick what it reads. Tick folders and assets in the side panel or the chat bar, a whole category in one click, or your personal folder beside them. Or type @: "@Pricing", "@cm-deal", "@All Wealth" for a whole category. The answer says what it read: "Read @Pricing".
• Tags, shaded. A tag shows a light highlight as you type and in the message sent: orange a folder, blue a category, green an asset.
• Talk as long as you want. Whisper large-v3 turbo writes what you say, through OpenRouter on the same key: punctuation and names included, in 27 languages or any with Detect, about $0.04 an hour of talking. The words arrive a piece at a time, about every 20 seconds, cut at a pause; a pause costs nothing. One strip over the bar shows who writes it, "Whisper · French", and for how long. Done keeps all of it, Cancel keeps nothing. The browser's free voice input stays one tap away in Settings.
• Follow-ups that hold. An answer remembers the concepts it read, so "and the second one?" finds what it named.
• Ask twice, pay once. The same question on unchanged folders returns the saved answer for 14 days, with no model call.
• It runs in the background. Open another chat, a folder or Drop: the answer lands in its own chat, and the inbox counts it.
• Errors say what to do. Try again resends the message in place. Send feedback mails the error, 5 an hour at most.
• Chats stay. Rename them and pin up to 5. The side panel shows 5, and See all opens the rest. Old ones clear after 30 days.

Folders

• A drop keeps everything. Every number, date and name of the source is checked against what was filed, and anything missing goes in word for word. The card says it: "All of it kept: 7,020 words, 145 numbers, dates and names, 23 topics, filed into 18 concepts".
• Categories. Wealth, Business, Life: each folder and each asset sits in one, or in none. Set them in one sheet, the grid next to +, with a suggestion from the model. A new version of an asset keeps its category.
• Public folders. Edit, Public: anyone adds the folder to Claude, ChatGPT or Cursor with its own address, and the home page lists it with its counts. It reads only; you alone feed it. A private folder never reaches the public connector.
• A whole category at once. Settings, Share folders: pick a workspace, or Public, then tap a category. Every folder of it is shared, or made public, in one tap. "2 of 5 public" says where it stands. A folder put in the category later waits for its own tap.
• One +. It asks first: Knowledge, a new folder, or Asset: files, a link, a blank doc, page or table. A file dropped on the list becomes an asset.
• A folder, open. Its scope and counts (concepts, sources, open clashes), then every concept, newest first. Chat in it, drop into it, rename it, merge it into another, or delete it with its Delete button. Delete first counts what goes: "Content holds 42 concepts and 12 sources, 9 of them only here".
• Every drop grows the graph. Each new idea links to its 6 nearest in meaning, in any language: needs, causes, supports, contradicts or example. A question reads what its ideas link to; a why, how or in-what-order question follows "causes" and "needs" one step further, up to 4 more. Folders sort themselves into topics.
• What follows. Two linked ideas from different folders can add up to a third. A drop writes up to 5, marked as Tasu's own conclusion.
• One-pager, above Drop. Folders, a category, up to 4 assets, or a question become a summary, a quiz, a deep dive or use cases, in 9 languages. Assets alone are read alone.
• Decisions, with a call ready. Clashes, concepts filed twice, titles not in English, empty positions and the same person on two cards sit in one list, grouped by folder. Each has a suggested call and its reason: "B holds. The later claim: September 2026 over February 2026". Accept one, or all in one tap.
• Audit. After 50 new sources, a folder is audited on its own the next time you open Tasu. The personal folder asks every 7 days.

Inbox

A bubble at the top right counts what waits for you: answers ready, questions on their way, a drop in progress, the weekly audit, the decisions, and what is still open with your people. A ring turns on it while something runs.

Settings

Model	GLM 5.3 Flash by default, for the whole workspace. Or a list of favourites: each morning the cheapest runs, when it is at least 10% cheaper. admin:addFavourite adds one more to the list.
Answers	Short reads the 12 best concepts, for about a third less a question. Long reads up to 30. Educational and Learning answers read in full either way.
App	Installs Tasu on the home screen or the dock. See Install as an app.
Appearance	System, Light or Dark, kept in this browser.
Languages	The language of answers (as you write, or English). Voice: Whisper (any length, any language) or the browser (free, 2 minutes). The mic language: 27, or Detect.
Cost	This month in dollars for each kind of work: folder chat, personal chat, assets, drops, upkeep, one-pagers and voice, and last month in all.
Export, Restore	Every folder as one markdown file, and read back in one step: what is missing returns, what is there stays.
Audit, Look, Passphrase, Share	Audits and clashes; a logo, an accent and a page colour; a new passphrase that signs others out; a folder shared with another workspace.

Install as an app

Tasu installs from the browser. The app stores play no part.

Device	Where to tap
iPhone, iPad	Safari, Share, then Add to Home Screen
Android	Chrome menu, then Install app
Mac	Safari, File, then Add to Dock
Windows, Mac, Linux	Chrome or Edge, the install icon in the address bar

Settings, App opens the browser's own prompt where there is one, and shows these steps everywhere else. Installed, Tasu opens in a window of its own, on the workspace used last, with the Tasu icon.

The personal folder alpha

A third kind of folder, next to subjects and experts. You talk to it, and it keeps what you said.

• It files your words. Each message becomes up to 3 notes, each dated and signed "You". A change of mind rewrites the note and keeps the old view with its date.
• Add memory. Paste what another assistant knows about you, or add a PDF, Word, text or JSON file. It is checked like a drop, and the result says "All of it kept".
• It answers as your twin. "I want Porto in 2027." It decides with your own rules and names them: "I'd pass. I never trade learning for 20% more salary."
• It calls your other folders. When one bears on what you said, it brings it up: "your Health folder says 3 to 5 g a day". Tick folders beside it, and it calls those.
• It fills your assets. Name one with @: "add what you know on @cm-deal". It reads the asset with your notes and your people's cards, and proposes the change to apply. The message is filed as notes too.
• A person is a folder. Overview, Facts, History a year at a time, Their people, and Raw notes word for word. A changed fact keeps the old one with the day it ended.
• Still open, in the inbox. The 3 oldest promises and follow-ups from your people, each with a tick. Add one by hand on a person's card. Comment under a line, and one send updates every line you commented.
• It interviews you. 335 quick questions in 17 chapters, life story first, each under 30 seconds. It skips what your notes already answer, and ends at 100%.
• Two lines under it in the side panel. Interview, with how complete your twin is, and Daily log, with today's count. A click starts that one in a new chat.
• A daily log. Daily log, in the side panel or next to Twin test, opens a chat of its own and asks about today, one question at a time: where you were, what you wanted, what blocked it, what changed, the result, the day around the work, the lesson and the why. A vague answer gets one follow-up for the number or the name. The change and the result always get one question more that digs in: what led to it, what it cost, what you do with it tomorrow. Each answer goes into your notes. Daily log reopens today's chat when there is one.
• Your twin, tested. 5 situations built from your notes. Your twin replies first, then you do, and the two are compared on the decision, the facts and the voice. Your replies go into your notes.
• Any language in, English kept. Write or talk in any language. Notes are kept in English, your own words beside each line.
• Private. Its own chat is its only reader.

Assets

Documents, tables, pages and web pages you keep. Any chat calls one with @, as it calls a folder.

• Add Word, PDF, text, Markdown, HTML, Excel or CSV files of up to 30MB, read in your browser. Or a link: the page is read and kept, up to 60,000 characters, and read again in one tap. Or a blank Doc, HTML page or Table.
• In the same list as the folders. A click ticks an asset for the chat, open shows it. Tick up to 4, here or in the chat bar, and every message reads them beside the folders ticked.
• Call it with @. "@invest-assets.csv what are the best 4 lines?" A table of up to 120,000 characters (about 300 rows of 30 columns) is read whole; a bigger one, up to 40,000 rows, by a filter and totals run over every row. A document of up to 60,000 characters (about 20 pages) is read whole; a longer one, up to 4,000 pages, by the 8 sections that bear on the question. Up to 4 assets and folders in one message.
• Nothing changes without you. A chat proposes a change with the words before and after, or a new asset with what it will hold. The answer keeps a line for each; the cards to decide open in a panel left of the chat, a sheet on a phone, gone once decided. Apply, Turn down, then Undo. Ask for a reply and a change in one message, and you get both. The personal chat proposes them too, from your notes.
• A guide for the AI. Up to 3,000 characters on what the asset is for, what each column holds and how a row is filled. Every chat that reads or changes the asset reads it first. Draft it for me writes one from the columns and first 12 rows.
• Versions. A file dropped under a name an asset holds becomes its new version. The one before is kept: Undo brings it back.
• Cost, on the default model (GLM 5.3 Flash, $0.035 a million tokens read, $0.90 written). A question on a 300-row table reads about 40,000 tokens: $0.0014, plus about $0.0005 for the answer. A proposed page or table of 3,000 tokens adds $0.003. Reading a 100-page PDF in costs about $0.005; a table or a link, nothing. Counted under Assets in Cost.
• See it. A document's pages, a table's grid, a page as it renders or as code. Download a document as Markdown, PDF or a web page, a table as Excel, CSV or Markdown.
• Projects from before. A line at the top of the list deletes them, or keeps them as assets, after a list of what each would become: every file with all it holds, and the Brief, where it stood and what it kept as one document, Notes of the project. Its last 4 chat exchanges go. Nothing moves before you confirm.

Workspaces

A workspace holds its own folders and chats, apart from every other.

	Octopus, Squidgy	The demo	Your own
Whose	The builder's, and someone's	Anyone's	Yours
Enter	Open a workspace, then its passphrase	One click, under the list	Made from the landing, then its passphrase
Model calls paid by	The deployment's key	A separate demo key	Your OpenRouter key
Limits	None	30 drops and 300 questions a month, shared	Your key's own
New folders, personal folders	Yes	No	Yes

• A workspace for someone, on your key. npx convex run admin:makeWorkspace "{name:'PandAAAHH',pass:'ABC12345'}" --prod makes an empty workspace behind that passphrase. Its owner changes the passphrase from Settings.
• Share a folder. One folder, two workspaces, one copy. A drop in either fills both, and a clash ruled in one is ruled in both. Shared with the demo, visitors can only ask it. A personal folder stays private.
• Your key stays yours. It is checked once with OpenRouter, kept in your browser, and sent only with the calls that need a model. Sign out forgets it.

Run your own

A Convex deployment holds the data. Any static host serves app/.

git clone https://github.com/jlasne/brain && cd brain && npm install
npx convex dev
npx convex env set OPENROUTER_API_KEY sk-or-... --prod
npx convex deploy
npx convex run admin:setPass '{"space":"octopus","pass":"at least 12 characters"}' --prod
npx convex run admin:makeDemo '{"copy":["health","content","social"]}' --prod

Point window.OCTOPUS_API in the pages under app/ at your deployment's .convex.site address. In Windows PowerShell, write each argument as "{space:'octopus',pass:'...'}".

Variable	What it does
OPENROUTER_API_KEY	Pays for model calls in the owner's workspaces. Required.
DEMO_OPENROUTER_API_KEY	A separate key for the demo, so its spend shows apart. Falls back to the one above.
DEMO_MONTHLY_DROPS	Drops the demo takes in 30 days, all visitors together. Default 30.
DEMO_MONTHLY_ASKS	Questions the demo answers in 30 days. Default 300. One address takes a tenth of each demo limit at most.
DEMO_MONTHLY_STEPS	Reads, plans and stores the demo makes in 30 days. Default 20 a drop.
SUPADATA_API_KEY	Fetches YouTube captions from a bare link. Without it, a video link asks for the transcript.
SUPADATA_MODE	auto transcribes videos that have no captions, at 2 credits a minute.
OTHERS_FETCH_DAY	Pages and transcripts fetched for every other workspace together, a day. Default 100, on top of 20 each.
RESEND_API_KEY, MAIL_FROM	Mail one-pagers from the owner's workspaces.
HOSTED_MAX_PRICE	The most a model may cost a workspace made with admin:makeWorkspace, in dollars a million tokens. Default 1.
PASS_PEPPER	Optional. A secret that keys the mark the landing's one field uses to skip doors. Set it once, before the first passphrase.
DIGEST_TO	Where the Monday digest goes: the week's sources, new concepts and open conflicts.
FEEDBACK_TO	Where Send feedback mails an error. Unset, it goes to DIGEST_TO.

digest:send '{"dry":true}' previews the digest.

A public folder. Its line on tasu.ai gives its address, tasu.ai/mcp/v1?b=<folder>. Paste it into Claude, ChatGPT or any MCP client: it reads that folder only, and the client's own model does the thinking.

Your workspace. Settings gives an MCP address carrying your key. It reads and feeds that workspace's folders. /doc has the steps.

Claude Code. Copy the skill folder under skill/ into ~/.claude/skills/. The folders are then plain markdown files, following the 41 rules.

• Passphrases are stored as PBKDF2-SHA256 hashes, 600,000 rounds, salted, one per workspace. An older hash is rewritten at its next sign-in. A new passphrase takes 12 characters.
• One address gets 8 guesses an hour at a door; all addresses together get 60. Each guess is counted before it is checked.
• A link is fetched only when its name points to a public address, checked at a public DNS service on every redirect. A page is read up to 2 MB.
• Every route checks the session before it reads a folder or calls a model.
• The demo runs on the default model within its monthly limits. Creating, renaming and connecting folders are closed to it.
• A personal folder is read by its own chat only. Every other reader loads the workspace without it.
• The server keeps only what was extracted: a visitor's OpenRouter key and raw transcripts stay out of it.
• Linking runs later on the deployment's key, so it stays off in the demo and in a workspace on its own key.
• The Word reader loads from cdnjs only if its bytes match the published hash.

Work on it

The app's source sits in web/chat/, one file a screen. npm run build joins it into app/chat.html, the one file the host serves. Commit both.

npm run build     # web/chat into app/chat.html
npm run check     # the build first, then 2025 checks
npm run deploy    # the server, to Convex

app/            the landing, the built app, the white paper, the install manifest and icons
web/chat/       the app's source: page.html, 10 stylesheets and 21 scripts, one a screen
convex/         the server: routes, store, drop, one-pager, conflicts, personal, assets
scripts/        the build and the checks, the app driven in a real browser included
docs/           the 41 rules (PROTOCOL.md), the server build (SERVER.md), the screenshots
skill/          the same behaviour, as a Claude Code skill
templates/      the markdown shape of a folder

npm run check runs 2025 checks: the build matches its source, the pages parse and bind, the install manifest and its icons, the MCP connector, the store and its workspaces, public folders, one-pagers, asking, linking, the personal folder, assets, and the app itself in a real browser at 9 screen sizes.

Built in September 2026 by Jeremy Lasne for The Build Games. MIT license.