AgentX MCP Server

Put any MCP-compatible AI agent under AgentX governance: identity, ABAC permission checks, audit trail and inter-agent messaging — without changing the agent's code.

How it works

Claude / any MCP agent ──stdio──▶ agentx-mcp ──HTTPS──▶ AgentX Rust API
                                                          │  ABAC policy engine
                                                          │  Audit log
                                                          └  Anomaly detection

The server authenticates with an agent API key (created in the AgentX dashboard → Agent → Credentials) and exposes five tools:

| Tool | What it does | |---|---| | agentx_whoami | Authenticate and return the agent's identity, status and rate limits | | agentx_check_permissions | List resources/actions allowed or denied by the agent's policies | | agentx_declare_action | Declare an action before doing it — policy decision (allow/deny) + audit record; supports dry_run | | agentx_send_message | Send a message/handoff to another agent via a channel | | agentx_receive_messages | Poll a channel for pending messages |

Setup

cd integrations/mcp-server && npm install

Claude Desktop / Claude Code configuration:

{
  "mcpServers": {
    "agentx": {
      "command": "node",
      "args": ["/path/to/agentx/integrations/mcp-server/index.js"],
      "env": {
        "AGENTX_API_URL": "http://localhost:8080",
        "AGENTX_API_KEY": "agx_..."
      }
    }
  }
}

Recommended agent prompt

Before any privileged operation (file write, API call, database access), call agentx_declare_action with the action and resource. Proceed only if allowed is true.

This gives you a complete, queryable audit trail of everything your agents attempted — allowed or denied — in the AgentX dashboard.