What the MCP server is
An MCP endpoint that exposes your own JW3 account — reads, proposals and execution — as tools any compatible AI client can call.
Aún no está en tu idioma — mostrando inglés.
The Model Context Protocol (MCP) is a standard way for AI applications to call tools on a server. JW3 runs one MCP server that exposes your account: balances, positions, diary, risk posture, market analysis, proposals and — if you grant it — live order execution.
Your client brings the model. JW3 brings the data and the order path. Neither side gets your keys.
Endpoint #
https://www.jw3.ai/mcp- Transport: Streamable HTTP, JSON-RPC 2.0 over
POST - Protocol version:
2025-06-18 - Authentication:
Authorization: Bearer <your token>on every request - Stateless: no session to keep alive, no socket to hold open
Requests without a valid token fail closed with -32001 unauthorized. There is no anonymous mode.
What it can do #
| Layer | Examples | Can it move money? |
|---|---|---|
| Read | Balances, positions, open orders, fills, realised PnL, diary, risk limits, news | No |
| Analyse | Multi-timeframe verdicts, prediction-market analysis, DeFi market data, portfolio insights | No |
| Propose | Trade setups, brackets, prediction trades, close-all plans | No — they file a card for your approval |
| Execute | Place an order, close a position, set exits, cancel an order | Yes, when the token carries trade:execute and your account allows it |
What it cannot do #
- See your keys. Wallet keys never reach the server; signing happens in your browser
or wallet. DeFi transaction builders return unsigned payloads for you to sign.
- Withdraw anything. No tool proposes or performs a withdrawal. Moving funds off an
exchange stays in the exchange's own interface.
- Bypass your limits. Every execution call re-checks the kill switch, autonomy level,
position caps, leverage caps and daily loss limits server-side.
- Act without an audit trail. Every call is attributed to a token and recorded.
Why a token instead of a password #
A token is scoped, revocable and identifiable. You can give a research client read-only access, a planning client propose rights, and an execution bot the ability to trade — each with its own token, each revocable on its own. See Connect your AI client and Permissions and safety.
When to use MCP instead of the CLI #
| You want | Use |
|---|---|
| A conversational assistant that can look things up | MCP |
| A scripted pipeline or cron job | CLI |
| A model that reasons about your book and suggests setups | MCP with propose scope |
| A deterministic guard process (stops, supervision) | CLI |
| Backtesting or batch analysis | CLI |
Both surfaces sit on the same account, the same limits and the same audit trail. You can use both with different tokens.
Checklist
- You know the difference between read, propose and execute tools
- You have an MCP client that supports remote HTTP servers
- You plan to start with a read-only token