Claude Tower

A fifty-floor building. Every floor is one automated Solana research desk. Its core decision path has fourteen seats including the CEO; the standing Regime and Review seats bring the permanent team to sixteen. Grok floors may add an alternate PM brain.

Floor 50 is the headquarters. Floors 1–49 are tenancies paid once in $CLAUDECO.

The hosted desk is research and paper accounting only. It never receives a private key, never signs a wallet transaction, and produces an unsigned order slip and a GMGN link. This repository also contains WALL-ST-E, an isolated polling executor that runs on the user's own Linux host. It defaults to paper mode; an explicitly armed, supervised live canary can sign from a dedicated local burner only after the durable journal, transaction validation, two-private-RPC, freshness, fee, slippage, and risk gates pass. The legacy webhook executor remains dry-run-only. Browser signing is absent and the old hosted RPC relay returns 410 Gone without contacting Solana.

---

The seats

Deterministic stages are plain code. Judgment stages are Claude. They alternate on purpose: code narrows the field cheaply, the model only reasons about what survived.

Stage	Seat	Kind	The one question it answers
0	Scout	code + model	What deserves attention today, and why now?
1	Screener	code	Does this clear the floor at all? (kills most candidates, costs nothing)
2	Forensics	model	Can this token be used against a holder, by design?
3	Liquidity	model	Can I get out, at size, at a price I'd accept?
4	Flow	model	Is the demand real, or manufactured?
5	Narrative	model + web	Is there a story, is it true, and am I early?
6	Red Team	model	Why does this trade lose money?
7	Risk	model + code	Choose a tier and thesis stop; code derives size and loss.
8	PM	model	Propose, watch, or pass — on what thesis?
9	Execution	model	The unsigned ticket: route, slicing, targets — the stop is Risk's, verbatim.
10	Compliance	code	Does this break a house rule? (veto, not advice)
11	CEO	model	Do I trust this desk, on this trade, today?
12	Scribe	code	Write it down so the desk can be graded later.

A Technical seat sat between Narrative and Red Team until 2026-09-08. It was retired on its own record — 31 calls and $0.67 in a live day, 0 kills, and no kill clause in its brief to make one; the four price-change windows it read are on the bundle for every other seat.

Two permanent seats sit outside that per-token sequence: Regime computes the broader SOL/BTC weather used by the evidence and portfolio gates, while Review grades each closed call and gives one transferable lesson to the next PM decision.

Codex is available as a separate Improvement Engineer, not a trading seat. It reviews the repository, exact prompt/test provenance, and aggregate house evaluation evidence in a read-only sandbox during a manually dispatched GitHub Actions run, then emits at most five structured proposals for human review. It has no route into Risk, PM, Execution, Compliance, CEO, call publication, wallets, or deploys.

Stages 2–6 run in parallel and blind to each other. Five analysts who read each other's notes produce one opinion wearing five hats; the desk needs five actual opinions.

The Red Team is not asked to be balanced — it is asked to destroy the idea. A proposal survives only if the PM can answer a structured, verified fatal fact with retained evidence. An answered refutation may reach the CEO at mechanically reduced size; unsupported fatal prose is retained as a wounded finding, not promoted into a hard veto.

The CEO is the only seat shown the desk's own historical record, because its question isn't "is this a good trade" — the PM already answered that — but "do I trust my desk today".

What a tenant actually rents

A floor has two complementary paths. House calls are copied to it at no model cost only when they clear that floor's deterministic mandate. A tenant can also dispatch a metered full-team workup against a mint; that run has its own journal and, if it clears the same publication gauntlet, can publish a call only to that floor. Settings, deliveries, runs, and floor-only calls remain tenant-scoped.

Runs are metered, not continuous, and the reason is arithmetic rather than caution. The paid seats are deliberately tiered: Scout defaults to Haiku; the evidence-shaped analysts, Risk and Execution to Sonnet; Red Team, PM and CEO to Opus. Effort and retained evidence make the cost variable, but every workup consumes metered model credit. Forty-nine unbounded desks cannot be funded by a one-time lease, so the product meters runs instead of promising unlimited compute.

So: a lease includes FREE_RUNS_WITH_LEASE runs, and further runs cost RUN_PRICE_CLAUDECO from the same $CLAUDECO credit balance the lease was paid from. The tenant points their team at a token; the core fourteen-seat path works it end to end; the tenant gets a brief and an unsigned ticket they sign themselves, hand to the optional local executor under its own gates, or decline.

The building

Page	Route	What it is
Website	/	The front door: what the desk is, who the sixteen permanent seats are, what a floor costs
The tower	/tower	All fifty floors in 3D. Click one to select it; the directory tracks it
A trading floor	/floor/:n	One desk's isometric office, with its agents at work

The 3D is built to the Markets & Makers house style, read out of that codebase rather than guessed at: a true-isometric orthographic camera (45° yaw, 35.264° elevation — that is atan(1/√2), not 35 and not 30), AgX tone mapping at exposure 1.12, its exact three-light rig, and its blocky 21-box citizens bound rigidly to a 15-bone skeleton.

Two house rules that matter, both load-bearing:

• The room has no front wall and no ceiling, and its side walls are 0.45 units high. That is not laziness — it is how that game solves camera occlusion, architecturally, so nothing ever has to be hidden or faded between the camera and a character.
• Merge, don't instance. Every static prop is baked into one geometry per finish and every avatar is a single skinned mesh. That took the floor from 1,045 draw calls to 99.

Deploying the website

The site is fully static — every page runs with no backend (the tower and the floors fall back to a scripted demo shift when there is no desk process to talk to). It deploys to GitHub Pages automatically:

1. Create the empty GitHub repo and git push -u origin main.
2. In the repo: Settings → Pages → Source: GitHub Actions (one time).
3. Every push to main then builds dist/ and publishes it. The workflow supplies the custom-domain SITE_URL and API base so link-preview images and API calls resolve.

Local equivalents:

npm run build                                 # local preview; executor copy buttons are disabled
INLINE_ASSETS=1 npm run build                 # local artifact preview with images inlined
EXECUTOR_COMMIT="$(git rev-parse HEAD)" npm run build  # pinned release build from a reviewed clean commit

CI supplies its exact GitHub commit automatically. A release build rejects malformed commit pins; an unpinned local preview shows no runnable executor install command.

A custom domain later is Settings → Pages → Custom domain, plus a CNAME at your DNS.

Deploying the API safely

The hosted API is described by render.yaml. Render installs from the lockfile and runs the isolated root test suite before starting a new revision:

npm ci
npm test

A failed test therefore fails the Render build instead of replacing the live process. Production SQLite state lives on the persistent disk at /var/data/claude-co.db; the ignored database in a local checkout is not production and must never be used as a test target. The test runner creates throwaway databases for this reason.

Executor webhook delivery is explicitly disabled in the production blueprint with EXECUTOR_WEBHOOKS_ENABLED=0. The authenticated polling feed remains available to local paper or explicitly armed canary executors, but the hosted server will not push entry or exit events to executor URLs. Unsigned transaction preparation is also disabled with DESK_PREPARE_TX=0. The retired browser RPC endpoint always returns 410 Gone and never contacts the upstream Solana RPC.

After deployment, verify /api/lease/config, /api/stats/overview, /api/heartbeat, and /api/improvements/status. /api/storage says what the database disk holds: the file, its write-ahead log, free space inside the file, the disk around it and the biggest tables by estimated rows (aggregates only). Bytes per table take a walk of every page, which froze the API for over a minute on the 1.4 GB file, so that runs only as its own process from the Render shell: node src/index.js storage. The same hourly job deletes insufficient_coverage decisions (fewer than three analyst seats returned; excluded from every measurement and never publishable) once they are a day old, with their forward marks and simulated outcomes, in small batches. With the Anthropic key disabled the research loop wrote ~19,000 of these a day; PENTHOUSE_ENABLED=0 stops the loop itself. The API resets the write-ahead log hourly and logs [storage] WARNING once the disk or the database passes 80% of the disk, naming the biggest tables. The tenant dashboard additionally reads /api/candidates/board for the immutable pre-decision top-five board, /api/callouts for exact author-wallet matches to confirmed pool-touching token inflows valued at the current market mark (not reconstructed purchase cost), and /api/floor/:n/executor/status for the floor owner's sanitized WALL-ST-E telemetry. Anonymous requests to the first two should be refused; the executor-status route must also refuse signed-in guests who do not own that floor. A ready improvement service reports bundleAuthConfigured: true, and its sourceCommit must equal the exact 40-character commit SHA deployed by Render before an improvement review is dispatched. The public status response is deliberately coarse; the review bundle itself is not public. A reachable HTTP server does not prove that the model seats can work. /api/heartbeat reports BLOCKED when recent Anthropic credit failures have not been followed, at least five minutes later, by a successful paid seat. Its providerCredit object exposes the failure and success timestamps, failure count, and recovery grace used for that decision. RUNNING is acceptable after a prior credit failure only when blocked is false and lastSuccessTs >= lastFailureTs + recoveryGraceMs.

Running it

cd ~/Downloads/claude-co && cp .env.example .env

Set ANTHROPIC_API_KEY in .env. Everything else has a working default, though a private RPC is strongly recommended — the public one rate-limits holder queries, which costs the Forensics seat real confidence.

An API key without available Anthropic credits is not operational. The HTTP process may remain reachable in that state, so use /api/heartbeat rather than transport health to decide whether paid research is available.

npm run doctor

npm run one -- DezXAZ8z7PnrnRJjz3wXBoRgixCa6xjnB7YaB1pPB263 --office

npm run desk -- --office

Command	What it does
npm run doctor	Config, RPC reachability, journal stats
npm run one -- <mint>	Full workup on one token
npm run desk	Scout the feeds, work up the shortlist
npm run watch -- 30	Same, every 30 minutes, with the floor open
npm run ledger	Every proposal the desk has made
npm run improvement:bundle	Print a local, content-addressed aggregate review bundle for diagnostics
npm test	Run the isolated root regression suite used to gate Render deploys
node src/index.js office	Serve the site, the tower and the floors
npm run build	Bundle standalone pages into dist/ with three.js inlined
npm run check	Run the full regression suite, then build the static site

Add --office to open the floor at http://localhost:4949. Append ?demo=1 to watch a scripted shift without spending anything.

Codex Improvement Engineer

The improvement service is deliberately split in two:

1. The live API exposes a coarse public status at /api/improvements/status. The full /api/improvements/review-bundle is protected by a dedicated bearer token and returns house-only aggregates, evaluation scorecards, prompt/policy hashes, and the exact test manifest. It contains no symbols, mints, workup prose, tenant rows, wallets, sessions, executor configuration, provider URLs, or credentials.
2. The manually dispatched Codex improvement review GitHub Actions workflow is the only supported worker environment. It verifies the bundle's sourceCommit against the exact checked-out GITHUB_SHA, verifies both manifests and the content digest, runs Codex with a read-only sandbox, approvals and agent networking disabled, and writes JSON and Markdown proposals. It never runs inside Render's trading process and never receives the production database or trading secrets.

To activate it, configure a high-entropy CODEX_REVIEW_TOKEN in Render and set the GitHub Actions secret of the same name to the exact same value. Also configure these repository Actions secrets:

• OPENAI_API_KEY — project-scoped credential used only by the isolated worker.
• CODEX_REVIEW_TOKEN — read-only bearer credential matching Render.
• CODEX_ARTIFACT_KEY — a separate random passphrase of at least 32 characters.

Redeploy Render, request /api/improvements/status, and confirm bundleAuthConfigured is true and sourceCommit is the exact commit SHA the workflow will check out. A missing Render token leaves the full endpoint unavailable; a missing or mismatched GitHub token cannot fetch it. Dispatch the workflow only after Render is on that exact commit, because the worker rejects stale or differently built bundles.

The workflow uploads only codex-improvement-review.tar.gz.gpg, encrypted with GPG and retained for seven days. Download it and decrypt it without putting the passphrase in shell history:

read -rs CODEX_ARTIFACT_KEY
printf '%s' "$CODEX_ARTIFACT_KEY" | gpg --batch --yes --pinentry-mode loopback \
  --passphrase-fd 0 --output codex-improvement-review.tar.gz \
  --decrypt codex-improvement-review.tar.gz.gpg
unset CODEX_ARTIFACT_KEY
mkdir codex-improvement-review
tar -xzf codex-improvement-review.tar.gz -C codex-improvement-review

Keep the artifact key outside the repository and rotate it periodically and immediately after suspected exposure. If old artifacts still matter, retain their prior key securely until their seven-day retention window expires. An artifact is advice, never a patch or approval: the service cannot auto-apply, commit, push, merge, deploy, or trade. Every proposal still requires a human-owned branch, full tests, review, and explicit merge. The worker also rejects policy-change proposals until there are at least 100 distinct, actually published assets with comparable 24-hour signals from the current decision manifest and at least 80% resolved-mark coverage.

The trading floor

A top-down view of the firm. Each figure is one agent, and they move because a pipeline stage actually started or finished — not on a timer. Analysts print their notes and carry them to the Red Team; the Red Team walks its verdict to Risk; Risk walks size to the PM; the PM stops at the threshold and slides proposals under the CEO's door. Rejects go in the shredder. Everything ends up in the journal.

The CEO's office is the one room no other agent enters.

The floor HUD has eight working destinations: Overview, Calls, WALL-ST-E, Team, Callouts, Activity, Performance, and Settings. Calls keeps the free-screened candidate board visibly separate from published and closed calls. Candidate rows have no execution action. WALL-ST-E is an owner-only observation/setup surface: it can show the local process's self-reported heartbeat and a confirmed read-only balance for the public burner address, but it has no remote-control or browser-signing path. Callouts only shows Pump.fun authors whose exact public wallet matches recent qualifying confirmed token inflows; unmatched conversation is omitted, scan coverage is explicit, and current-mark value is never relabelled as the wallet's original purchase consideration.

How an order reaches you

GMGN sits behind Cloudflare bot protection, so this desk cannot reach it server-side and does not try. Instead a published call becomes:

1. an order slip in reports/ — ruling, size, stop, conditions, and what the desk got wrong;
2. a GMGN deep link you open in your own browser, where your wallet lives;
3. optionally, a real unsigned Jupiter transaction, if you set DESK_WALLET_PUBKEY.

DESK_WALLET_PUBKEY is a public key. The hosted desk never reads, requests, stores, or accepts a private key or seed phrase. WALL-ST-E loads a dedicated burner only on the user's machine; paper mode is the default, and the sole live-capable path is the local polling canary documented in executor/README.md. The legacy webhook adapter remains dry-run-only.

What it costs

Model defaults are cost-aware: Haiku scouts, reads the book for the Liquidity seat and writes the Execution ticket; Sonnet handles the rest of the structured evidence work; Opus is reserved for Red Team, PM and CEO judgment. The running total is printed each cycle. The screener is the main cost control: it is pure code and rejects most candidates before a model call. To spend less, lower DESK_MAX_CANDIDATES or override an individual DESK_MODEL_* setting; deterministic seats have no model cost. DESK_EFFORT_REDTEAM (low/medium/high/xhigh/max) is an A/B handle for the Red Team's thinking effort only — the default stays high, because its verdict feeds a safety gate.

Known limits

• fdv_propped mis-fires on CEX-listed majors. The ratio compares FDV to DEX liquidity, so a token whose real depth is on centralised venues looks thin. It is the right test for the DEX-native tokens this desk actually scouts; BONK trips it.
• There is no candle history on the bundle — four price-change windows only. The seat that used to read them alone (Technical) is retired; no other seat is asked to invent chart levels from them.
• Holder concentration is often unavailable on the public RPC, and the desk reports that rather than estimating it.
• Nothing here has an edge until you have graded it. Read the first weeks of the journal as a backtest you are watching forward.
• The live executor is an experimental local canary, not evidence of an edge. It is Metis exact-in/classic-SPL-only, code-bounded with locally acknowledged cap changes, and requires operator-owned infrastructure; browser and webhook signing remain disabled. Dry runs, simulations, and successful fills do not establish positive expectancy.

The charter

DESK.md is the firm's constitution, and it is injected verbatim into every agent's system prompt — so the rules bind the agents in-band, not just the reader.