prova-mcp

Verifiable AI inference for any MCP-compatible environment. Drop this MCP server into Claude Desktop, Cursor, VS Code, or any framework that supports the Model Context Protocol — every inference runs in a TEE, is hardware-attested, and verified on Solana.

Agents are making financial decisions, managing treasuries, and executing trades — with no proof of what happened inside the model. This MCP server fixes that. Every call becomes an un-fakeable, on-chain receipt.

Install

npm install prova-mcp

Configure

Get a Prova API key at prova.chat/dashboard.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "prova": {
      "command": "npx",
      "args": ["prova-mcp"],
      "env": {
        "PROVA_API_KEY": "prova_sk_your_key_here"
      }
    }
  }
}

Cursor

Create .cursor/mcp.json in your project:

{
  "mcpServers": {
    "prova": {
      "command": "npx",
      "args": ["prova-mcp"],
      "env": {
        "PROVA_API_KEY": "prova_sk_your_key_here"
      }
    }
  }
}

Claude Code (CLI)

claude mcp add prova-mcp -- npx prova-mcp
export PROVA_API_KEY=prova_sk_your_key_here

VS Code

Add to your MCP extension settings:

{
  "mcp.servers": {
    "prova": {
      "command": "npx",
      "args": ["prova-mcp"],
      "env": { "PROVA_API_KEY": "prova_sk_your_key_here" }
    }
  }
}

Available tools

Once connected, your AI assistant can call:

| Tool | Input | Returns | |------|-------|---------| | submit_proof | prompt, optional model | Response + Solscan link | | verify_proof | optional receipt_id | Full attestation result | | check_model | model name | TEE status + provider | | list_models | — | All 16 models with TEE details | | get_stats | — | Session proofs, anchored count |

What gets proven on-chain

• Which model ran
• Which TEE hardware signed it (Intel TDX or NVIDIA H100)
• Hash of the input (req_hash)
• Hash of the output (res_hash)
• ECDSA signature from the enclave
• Timestamp, provider
• verifiedSignature / verifiedRegistry — both enforced by Solana

Verify any proof at prova.chat/verify, or browse the on-chain verifier: D4fwqE74azXzC6euWAmDoH6Up1gZEh725odUuZHcCqEB

Supported models (16)

All models run inside a TEE with Intel TDX attestation:

• deepseek/deepseek-v4-flash (default)
• deepseek/deepseek-v4-pro
• deepseek/deepseek-r1
• deepseek/deepseek-r1-0528
• qwen/qwen3-235b-a22b
• qwen/qwen3-30b-a3b
• qwen/qwen3-32b
• qwen/qwq-32b
• thudm/glm-4-32b
• meta-llama/llama-4-scout
• meta-llama/llama-4-maverick
• meta-llama/llama-3.3-70b-instruct
• google/gemma-3-27b-it
• mistralai/mistral-small-3.1-24b-instruct
• nvidia/llama-3.1-nemotron-ultra-253b-v1 (NVIDIA H100 TEE)
• heretic.ai/hermes-3-llama-3.1-8b

Works with

• Claude Desktop + Claude Code
• Cursor
• VS Code (via MCP extension)
• ChatGPT (via MCP plugin)
• Any MCP-compatible client

Links

• Website: prova.chat
• Docs: prova.chat/docs/mcp
• Verify a proof: prova.chat/verify
• Get an API key: prova.chat/dashboard

License

MIT © Prova