@prova-ai/mcp-server

Model Context Protocol (MCP) server that wraps any MCP client in an EU AI Act Annex IV compliance envelope.

Install once, add three lines to your MCP config, and every tool call flows through the PROVA classifier + oversight gates + immutable audit log.

Install

npm install -g @prova-ai/mcp-server

Claude Desktop config

Open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the Windows equivalent, and add:

{
  "mcpServers": {
    "prova": {
      "command": "npx",
      "args": ["-y", "@prova-ai/mcp-server"],
      "env": {
        "PROVA_AGENT_ID": "claude-desktop-prod",
        "PROVA_RISK_PROFILE": "high-risk",
        "PROVA_ENDPOINT": "https://api.getprova.dev/api/ingest",
        "PROVA_API_KEY": "pk_live_..."
      }
    }
  }
}

Restart Claude Desktop. The prova_audit, prova_classify, and prova_status tools appear in the tool menu.

Environment

| Variable | Required | Description | | ------------------------ | -------- | ----------------------------------------------------------------- | | PROVA_AGENT_ID | ✓ | Registered agent id in the PROVA dashboard | | PROVA_RISK_PROFILE | | high-risk · medium-risk · low-risk (default: medium-risk) | | PROVA_ENDPOINT | | PROVA cloud ingest URL — if omitted, audit records are local only | | PROVA_API_KEY | | Bearer token for PROVA_ENDPOINT | | PROVA_AUDIT_PATH | | JSONL path for local audit log (default ./.prova-audit.jsonl) | | PROVA_OVERSIGHT_GATES | | Comma-separated gate names for Art. 14 escalation |

Tools

• prova_audit — run an audit envelope over a message (+ optional output). Returns {status, auditId, riskScore, annexIIICategory, blocked}.
• prova_classify — Annex III classification + risk score, no persistence.
• prova_status — diagnostic: live config and whether cloud ingest is wired.

License

MIT · © PROVA AI — getprova.dev