prova-mcp

Let any AI agent verify its own reasoning before answering — and kernel-check the proof on its own machine.

prova-mcp is a Model Context Protocol server that exposes the Prova reasoning verifier as tools any MCP client can call. Drop it into Claude Code, Cursor, Windsurf, Zed, or ChatGPT desktop and the host model gains a self-verification loop:

1. The agent drafts a multi-step argument.

2. It calls verify_reasoning on the draft. Prova returns a verdict (VALID / INVALID), a confidence score, and — for INVALID — the exact step that's broken.

3. If VALID, the agent calls kernel_check_proof on the emitted Lean 4 proof. The local Lean kernel either accepts every step or rejects the proof. There is no third option, and Prova is not in that loop — it's purely a property of the proof and the kernel on your machine.

Result: agents that catch their own circular arguments, contradictions, and unsupported leaps before they reach a user — with a tamper-evident certificate as audit trail.

Tools

It also exposes each certificate as a resource at prova://certificate/{id} for clients that consume MCP resources.

Install

pip install prova-mcp

Optional but recommended — install Lean 4 so kernel_check_proof actually runs:

curl https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh -sSf | sh

Configure

Set environment variables wherever your MCP client launches the server:

Wire it up

Claude Code

claude mcp add prova prova-mcp -e PROVA_API_KEY=sk_live_...

Cursor / Windsurf / Zed / ChatGPT desktop

Add to your client's MCP config (typically mcp.json or claude_desktop_config.json):

{
  "mcpServers": {
    "prova": {
      "command": "prova-mcp",
      "env": {
        "PROVA_API_KEY": "sk_live_..."
      }
    }
  }
}

A copy of this config is shipped at examples/claude_desktop_config.json.

How an agent uses it

A good system prompt nudge:

Before producing a multi-step argument, call prova.verify_reasoning on your draft. If the verdict is INVALID, repair the failing step and re-verify. If VALID, attach the certificate URL to your answer.

That single line turns Prova into a default reflex for the model — every reasoning answer ships with a verifiable certificate, and broken arguments are caught before the user ever sees them.

Why this exists

Verifiers are only useful if they sit where the reasoning happens. MCP is now the universal bridge between LLM clients and external tools — so shipping Prova as an MCP server makes formal reasoning verification the easiest thing to add to any agent stack on the planet. Install one package, set one key, get a tamper-evident proof of every important argument your agent makes.

Related

• Prova — the reasoning verifier itself; Lean 4 proofs, certificates, and the public API this server wraps.

• insinuateai/prova-action — sibling GitHub Action (uses: insinuateai/prova-action@v1) that gates CI on reasoning validity.

• Verify independently — kernel-check any Prova proof on your own machine in five minutes.

Changelog

0.1.0 — first release

• 5 tools: verify_reasoning, get_certificate, download_lean_proof, kernel_check_proof, verify_and_kernel_check.

• Resource: prova://certificate/{id}.

• Stdio transport; works with any MCP-compliant client.

• Configurable via PROVA_API_KEY, PROVA_API_BASE_URL, PROVA_LEAN_BIN, PROVA_DEFAULT_RETAIN.

License

MIT. See LICENSE.