nexus-al-mcp — Accountability Ledger MCP Proxy

Sits between your AI agent and any MCP server. Every tool call is intercepted, identity-verified, and logged as a tamper-evident compliance record.

Satisfies EU AI Act Art. 12 (logging) and GDPR Art. 30 (records of processing).

How it works

Agent → nexus-al-mcp proxy → upstream MCP server
                 ↓
         Rust API (/api/acm/*)
                 ↓
         PostgreSQL (tool_call_events, hash-chained)

Every tool call is logged before forwarding. Fail-closed: if logging fails, the call is blocked.

Quick start (dev mode — 30 minutes)

1. Get your credentials

Sign up at veridion-nexus.eu → register your agent → copy your Agent ID and Service Token.

2. Install

npx nexus-al-mcp

3. Configure Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "nexus-al": {
      "command": "npx",
      "args": ["-y", "nexus-al-mcp"],
      "env": {
        "AL_API_BASE_URL": "https://api.veridion-nexus.eu",
        "AL_SERVICE_TOKEN": "your_service_token",
        "AL_AUTH_MODE": "dev_bypass",
        "AL_DEV_CLIENT_ID": "agt_your_agent_id",
        "UPSTREAM_MCP_MODE": "stdio",
        "UPSTREAM_MCP_COMMAND": "npx -y your-upstream-mcp-server"
      }
    }
  }
}

4. View logs

Open app.veridion-nexus.eu → ACM Overview → see every tool call logged.

Environment variables

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | AL_API_BASE_URL | Yes | http://localhost:8080 | Rust API URL | | AL_SERVICE_TOKEN | Yes | — | Service token for /api/acm/* | | AL_AUTH_MODE | No | jwks | dev_bypass or jwks | | AL_DEV_CLIENT_ID | dev_bypass | — | Fixed agent client_id for testing | | UPSTREAM_MCP_MODE | No | stdio | stdio or sse | | UPSTREAM_MCP_COMMAND | stdio | — | Command to launch upstream MCP server | | UPSTREAM_MCP_ARGS | No | [] | JSON array of args for upstream command | | UPSTREAM_MCP_URL | sse | — | SSE endpoint for upstream server | | AL_ORIGIN_COUNTRY | No | DE | Origin country for transfer records | | AL_EEA_EXTRA_COUNTRIES | No | — | Comma-separated extra EEA codes |

Production (OAuth 2.1)

For production, set AL_AUTH_MODE=jwks and configure:

• AL_OAUTH_ISSUER — your authorization server URL
• AL_OAUTH_AUDIENCE — expected audience claim
• AL_JWKS_URI — JWKS endpoint (defaults to {issuer}/.well-known/jwks.json)

Every agent must have a registered oauth_client_id in the agents table.

What gets logged

Each tool call creates a ToolCallEvent with:

• Agent identity (from OAuth token, not self-reported)
• Tool name and input field names (not values — data minimisation)
• Context trust level (trusted / degraded / untrusted)
• SHA-256 hash chain (tamper-evident)
• OTel trace_id for delegation chain (when W3C traceparent header present)

Optional side-records:

• DataTransferRecord — when tool call involves non-EEA transfer
• HumanOversightRecord — when degraded trust + high-risk AI system

License

UNLICENSED — proprietary.