mcp-proofjson

A local MCP server (stdio) that lets AI agents produce and verify structured JSON outputs — with explicit scope, checks performed, checks not performed and human-confirmation requirements, instead of opaque yes/no answers.

The first live pack focuses on invoice assessment before payment: given supplied invoice data, context and policy, it returns allow / review / block with reasons and confidence. Works in Claude Desktop, Cursor and other MCP clients.

Scope (v1): supplied_data_only. It assesses the supplied data and configured policy. It does not independently verify supplier IBAN ownership, sanctions status, bank account ownership or external payment history unless external trust adapters are configured. allow is not a guarantee the invoice is safe.

Local stdio only — there is no public remote MCP until auth, rate limits, audit logs, terms and privacy are finalised.

Tools (exactly three)

• assess_invoice_before_payment — returns allow / review / block with reasons, confidence, scope, checks_performed, checks_not_performed, human_confirmation_required.
• verify_proofjson_proof — verify a proof for structure, hash consistency and signature validity. Does NOT verify the real-world truth of the underlying claim, invoice, supplier identity, IBAN ownership or sanctions status.
• list_proofjson_packs — lists the live Pack Tasks.

This server does not expose arbitrary fetch, a browser, a shell, the filesystem, a generic HTTP proxy, a generic run endpoint, or any internal scoring.

Install & build

npm install
npm run build

Use in Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "proofjson": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/mcp-proofjson/dist/index.js"]
    }
  }
}

Use in Cursor

Add to .cursor/mcp.json (or Cursor Settings → MCP):

{ "mcpServers": { "proofjson": { "command": "node", "args": ["/ABSOLUTE/PATH/mcp-proofjson/dist/index.js"] } } }

Config

• PROOFJSON_BASE (env) — defaults to https://proofjson.org.

Example

Ask your agent: "Assess this invoice before payment. Supplier ACME, amount 4250 EUR, invoice IBAN PT50…789, expected IBAN PT50…154." → the agent calls assess_invoice_before_payment → ProofJSON returns review because the IBAN differs from the expected IBAN supplied. The agent explains the result and recommends human confirmation.