Chp Mcp

Local MCP server that gates AI agent spending and tool calls with budget caps, human-approval thresholds and HMAC-signed receipts plus a chained audit ledger.

What you can do

• Evaluate spend requests against policy caps, confidence floors and human-approval thresholds, returning locked, HITL-required or blocked outcomes
• Record human approvals or denials and mint signed approval receipts bound to actor, tool, resource, arguments hash and expiry
• Check tool calls against allowlists and receipts, denying on changed arguments, host-bound overrides, expiry, replay or bad signatures
• Request authorization for scoped finance tools and run synthetic equity, treasury and rebalance operations only with a valid receipt
• Inspect the chained deny/authorize/execute audit ledger and compute canonical content hashes

When to use it

Use it when an agent can trigger spending or sensitive tool calls and you want a fail-closed gate that requires a human-signed receipt for each specific call, with a durable audit trail of refusals. It suits local Cursor or Claude Code setups that need more than a tool-name allowlist.

Requirements

Runs locally over stdio via npx or a global npm install of the package; Node.js is needed. Set CHP_RECEIPT_KEY (or AUDIT_LEDGER_KEY) for signing, and optionally CHP_AUDIT_LEDGER to a JSONL path or :memory:. No accounts or remote endpoints are declared.