1Claw Muse Connector

The 1Claw connector for Meta Muse. Muse users get their AI agents' approval queue, wallets and activity inside Muse — "is anything waiting on me?", "approve that", "what did my agents do today?" — while 1Claw keeps every credential, guardrail and audit trail exactly where it was.

Muse builds Custom Connectors by retrieving a service's API information and storing the user's credential in its Secure Credentials Store. This package is that API: a small, plain-spoken REST surface (/openapi.json, /llms.txt) in front of the 1Claw Platform API, plus the link flow that turns a signed-in 1Claw user into a scoped, revocable connector token.

Muse ──(mcn_ token)──▶ muse.1claw.co ──(plt_ key + connection id)──▶ api.1claw.co
                      this service                                   1Claw vault

What a Muse user can do

Ask Muse	Route	Notes
"Is anything waiting on me?"	GET /v1/approvals	Both queues (consensus + agent action approvals), one shape
"Approve that" / "No, reject it"	POST /v1/approvals/{id}/decide	The only write. Consensus votes echo the payload_hash the connector read, never one Muse supplies
"How much is in my agents' wallets?"	GET /v1/portfolio, GET /v1/balances	Only wallets on agents the user granted to Muse
"What did my agents do today?"	GET /v1/activity	Delegation log for the connection
"What jobs do they run?"	GET /v1/automations	Read only
"Which account is this?"	GET /v1/me	Connection + granted resources

Not possible through this connector, by design: reading secrets, moving funds, creating or changing agents or policies. Those stay in the 1Claw dashboard.

How a user connects (once)

1. Signed in to 1Claw, go to Settings → Connected Apps → Connect Muse (1claw.co/connect/muse). The page calls POST /v1/link on this service with the user's session; the service asks the vault who they are and creates (or finds) their connection on the 1Claw-owned Muse platform app. An existing 1Claw user goes through the standard "Sign in with 1Claw" link consent once.
2. The user chooses which agents Muse may see (the connected-apps grant flow — a connection with no grants sees nothing).
3. The page shows the connector token (mcn_…) once. The user pastes it into Muse when Muse asks for the service's API credentials, with base URL https://muse.1claw.co.
4. Revoke any time: Settings → Connected apps → Muse → Disconnect. Every request re-checks the connection, so a revoked token stops working immediately.

Security model

• The plt_ key never leaves this service. Muse holds an HMAC-signed connector token that names one platform connection and nothing else. Its mcn_ prefix is distinct from every vault credential (ocv_, 1ck_, plt_), so the vault will never accept it and a leak is identifiable.
• Stateless. No database: the platform connection is the state. Revocation = disconnecting the app; tokens expire after a year; rotating MUSE_CONNECTOR_SECRET invalidates all of them.
• Scoped by 1Claw, not by us. Every call is a Platform API call resolved through the vault's resolve_connection_for_platform gate, so the user's grants, the agents' guardrails, the sanctions screen and the audit hash chain all apply unchanged.
• Decisions bind to what was shown. A consensus vote must echo the pending approval's payload_hash; the connector reads it fresh and ignores any hash in the request body.
• /v1/link accepts only browser origins on the 1Claw dashboard allowlist and only with a 1Claw user session (a connector token is refused there); it never mints a token for an email the caller merely claims. The dashboard's /api/connect/muse route that fronts it is same-origin only.
• Ceilings in the service: 120 requests/min per connector token, 60/min per client address on /v1/* (bad tokens included, so the vault is never a token oracle), 10/min per address on /v1/link, 16 KB request bodies. Cache-Control: no-store, nosniff, X-Frame-Options: DENY on every response.
• Hostname: muse.1claw.co is served through the 1Claw edge (Vercel) proxy to Cloud Run, the same path as mcp.1claw.co; HSTS (includeSubDomains; preload) comes from the apex. The proxy forwards neither the dashboard session cookie nor a cookie-derived bearer to this service.

Running it

npm install && npm run build
MUSE_CONNECTOR_SECRET='<≥32 random chars>' \
ONECLAW_PLATFORM_API_KEY='plt_…' ONECLAW_PLATFORM_APP_ID='<uuid of the Muse platform app>' \
npm start          # listens on :8787

Env	Purpose
MUSE_CONNECTOR_SECRET	HMAC secret for connector tokens (≥32 chars). Rotate to revoke everything.
ONECLAW_PLATFORM_API_KEY	plt_ key of the 1Claw-owned "Muse" platform app
ONECLAW_PLATFORM_APP_ID	That app's id
ONECLAW_BASE_URL	default https://api.1claw.co
MUSE_CONNECTOR_PUBLIC_URL	default https://muse.1claw.co (OpenAPI servers)
ONECLAW_DASHBOARD_URL	default https://1claw.co
MUSE_CONNECTOR_LINK_ORIGINS	comma list; default the 1Claw dashboard hosts
PORT	default 8787

Docker: docker build -t muse-connector . && docker run -p 8787:8787 --env-file .env muse-connector.

Tests

• npm test — unit tests with a fake 1Claw port (tokens, the gate, queue merging, decide routing, link flow, CORS).
• npm run e2e:prod — creates a throwaway platform app on the SRE test account, provisions a user, starts the connector in-process and exercises every route end to end against api.1claw.co. Needs ONECLAW_TEST_EMAIL / ONECLAW_TEST_PASSWORD. Cleans up after itself.

Submitting to Muse

See connector/SUBMISSION.md for the directory listing text, the security and legal answers, and the end-to-end test plan Meta's review asks for.

License

Apache-2.0 — see LICENSE and NOTICE.