JAWZConnect · MCP

Connect Jawz

Setup and reviewer documentation for the Jawz MCP connector.

Jawz is a hosted Model Context Protocol (MCP) server. You add it once to an MCP-capable AI assistant (such as Claude or ChatGPT), and your assistant can then run the Jawz Loop framework and pull a live market read while you stay in your own AI app. Jawz is a connector your assistant calls; it is never the AI model itself.

Reviewers and the impatient: there is an unedited screen recording of a full session — five questions inside ChatGPT, no account and no key — at /demo.

Endpoint

The MCP server is served over streamable HTTP at:

  • https://jawz.ai/api/mcp

Connection is authenticated with OAuth 2.0. The authorization-server and protected-resource metadata are published at the standard discovery paths (/.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource), so a compliant MCP client discovers the sign-in flow automatically — you do not paste an API key.

Connect from Claude

  1. Open Settings → Connectors (or Custom connectors) in Claude.
  2. Choose Add custom connector and paste the endpoint https://jawz.ai/api/mcp.
  3. Complete the sign-in (OAuth) prompt in the window that opens. Once connected, the Jawz tools appear in your assistant.
  4. Ask your assistant to “list the Jawz loops”, “run the See the World chapter”, or “what’s this week’s Jawz read?” (the weekly World Brief, via get_world_brief) to confirm it works. Asking “what can Jawz do?” calls get_started, which returns first prompts and a live market read; the full walkthrough is in the Loop guide.

Connect from ChatGPT

In ChatGPT’s connector / apps settings, add a custom MCP connector pointing at https://jawz.ai/api/mcp and complete the same OAuth sign-in. Jawz is tested with both Claude and ChatGPT as MCP clients.

Do I need to sign in?

For reading: no. The market and macro reads, the Loop chapters and modes, the World Brief, and get_started — 22 tools — work with no token, no email, and no registration. Call the endpoint with no Authorization header at all. Anonymous use is rate-limited (300 calls per day, and 25 per day for get_prices / get_etf_profile, which spend third-party data quota), and counted only through daily-rotating hashes — never identified.

An email earns three things: your loop selection persists across sessions, you can submit feedback that shapes the Loop, and your rate ceilings rise. That is the whole trade — no password, no verification step, and Jawz stores only the email, your loop selection, and usage counts. Never your portfolio.

Server-side and headless agents

For the read tools you need none of this — call with no Authorization header and they answer. The flow below is only for the identity tools (loop selection, feedback), and it works without a browser as a supported contract, not a workaround. Four HTTP calls; registration is automatic (Dynamic Client Registration), so nothing is provisioned by hand.

  1. Register a client. POST /oauth/register with JSON {"redirect_uris":["http://localhost/cb"],"client_name":"my-agent"} → returns client_id. Register once; clients live for a year.
  2. PKCE. Generate a random verifier and its S256 challenge.
  3. Authorize. POST /oauth/callback (form-encoded): email, client_id, redirect_uri, state, code_challenge, code_challenge_method=S256, response_type=code → returns JSON {"redirect":"...?code=..."}. Extract the code from that URL. The email is the identity — no password, no verification step.
  4. Exchange. POST /oauth/token (form-encoded): grant_type=authorization_code, code, redirect_uri, client_id, code_verifier → access_token + refresh_token. Send Authorization: Bearer <access_token> to /api/mcp.

Token lifetimes. Access tokens last 24 hours (expires_in is authoritative). Refresh tokens last 30 days and rotate on every use: each refresh consumes the old token and returns a new one, so persist the new refresh_token from every response. A 10-minute grace window makes refresh retries idempotent — if your agent loses the response, retrying with the old token returns the same new pair instead of locking you out.

Recovery. If a refresh fails with invalid_grant, re-run the four calls with the same email. Identity and loop selection are keyed to the email, so nothing is lost.

What the tools do

The connector exposes two kinds of tools. Together they let your AI run a disciplined investing process rather than just fetch numbers:

  • The Jawz Loop framework — list and run the four chapters (See the World, Understand the Book, Decide, Observe & Refine) and their modes, and submit feedback on a run.
  • The market and macro read — prices for the symbols you ask about, plus a macro regime read, financial conditions, growth and inflation indicators, a global-liquidity read, and an event calendar, sourced from public data (such as the U.S. Federal Reserve’s FRED) and market-data providers.

Data handling

Jawz does not store your portfolio. When a loop runs with context that may describe your holdings, only a one-way hash of that text is kept — the raw context is never stored. Price lookups send only the ticker symbol (not your identity or quantities) to the market-data provider. The full detail, including what is and is not collected, is in the Privacy Policy; the terms of use are at /terms.

For a local-first companion that keeps your investment memory in plain files on your own machine — with no servers, accounts, or telemetry — see Rumo.

Support

Questions, access issues, or reviewer inquiries: hello@jawz.ai.