Skip to main content
Chance ships a local MCP server so any MCP-compatible desktop/IDE agent — Claude Desktop, Claude Code, Cursor, Windsurf — can verify a proposed action against its mandate before it executes, and get a signed, independently-verifiable verdict back. It uses the same API key and credits as the HTTP API.
Using a web or mobile assistant (Claude.ai, ChatGPT, Claude mobile), or one of the AI platform APIs? Those can’t launch a local process — use the hosted Connectors instead (one-click OAuth, no install).
Light integration. The MCP gives your agent the verdict. You still handle execution and decide to honor it — fail-closed: treat anything other than ALLOW as do-not-execute. For enforcement a failed check can’t ignore, give your agent an escrow wallet — it can only move funds through the harness.

Setup

1

Get an API key

Sign in at harness.chance.cc/login — 50 free credits, no card. In the dashboard, click Create API key and copy it.
2

Add the server to your MCP client

Settings → Developer → Edit Config, paste the block into claude_desktop_config.json, and restart.
Their MCP settings use the same mcpServers shape — paste the block into the MCP config and reload.
3

Use it

Restart your client and the verify_intent tool is available. Prompt your agent to call it before any consequential action and to honor the verdict.

Environment

Tools

tool
Verify a proposed action against the stated intent/mandate. Returns a verdict — ALLOW / BLOCK / ESCALATE — with a reason and a provable receipt (attested judge signature, hash-chained transcript, onchain anchor when configured). Costs one credit.Inputs: intent (string), action (string), context (object, optional).The verifier is platform-agnostic today — it works across prediction markets, perps/DeFi, and API venues (Hyperliquid, Kalshi) alike. Per-venue verification with protocol-specific knowledge bases is coming.
tool
Fetch a prior verification and its proof by id (UUID) or requestHash (0x…). Read-only, no credits.

Wallets your agent can spend from

The tools above verify an action your agent will execute itself. These go further: the wallet is held by the harness, so a BLOCK is never handed to a signer. See Escrow wallets.
tool
Provision a wallet on Base, Solana or Starknet. Inputs: name, chain, mode (autonomous executes on a passing verdict; safe waits for you to confirm), mandate, and optional limits — per-action and daily dollar caps, an address allowlist, and a token-contract allowlist.
tool
Propose a transfer: wallet, asset, amount, to, chain. Simulated, judged, then settled or held.
tool
Propose a swap. Chance fetches the quote itself, so the route that gets verified is the route that runs.
tool
Pay an HTTP resource that charges via x402. Chance requests the URL, reads the payment terms from the 402 response itself rather than from your agent, and verifies the exact amount, token and recipient. Use it for paid APIs and per-call data — and for topping your own credits up.
tool
Propose a raw transaction your agent built for any protocol.
tool
Read wallet balances, list wallets, and check the status, verdict and receipt of anything proposed. Read-only, no credits.
tool
Buy more verification credits over x402, without leaving the conversation.

Example

Prompt: “Before you place any Polymarket order, call verify_intent with my mandate as the intent and the order as the action. Don’t execute unless the verdict is ALLOW.”
The package is open source on npm as chance-mcp.