Cogtrix MCP server integration

Cogtrix is a first-class Model Context Protocol (MCP) client. You can plug any MCP-compliant server into Cogtrix and the agent treats its tools as native tools — same registry, same confirmation system, same on-demand loading.

What MCP is

MCP is Anthropic’s open protocol for exposing external capabilities to an LLM. An MCP server typically wraps an API, a database, a filesystem, or a domain-specific tool, and presents a standard JSON-RPC surface that any MCP client can call.

If you’ve configured an MCP server for Claude Desktop or Cline, the same config works for Cogtrix.

Connecting a server

Add a mcp_servers block to your .cogtrix.yml:

mcp_servers:
  filesystem:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]
    transport: stdio

  github:
    command: "npx"
    args: ["-y", "@modelcontextprotocol/server-github"]
    transport: stdio
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: ${GITHUB_TOKEN}

On startup Cogtrix spawns each server as a subprocess, handshakes over stdio, and registers every tool the server exposes.

Inspecting the connection

You> /mcp

Lists each configured server with its connection status, transport, and the tools it contributed. To reconnect a single server:

You> /mcp restart github

Tool naming

MCP tools are exposed in Cogtrix’s registry with the server name as a prefix to avoid collisions:

filesystem:read_file
github:create_issue

You can reference them in /tools load, in --activate-tools, and they appear in /tools output alongside built-in tools.

Transports

For remote servers, replace command/args with url: and the appropriate transport.

Limitations

• MCP resources and prompts are not yet routed through Cogtrix’s memory layer; only tools are auto-registered.
• Tool argument schemas larger than ~50 fields can slow down the model’s planning step. Trim aggressive servers if you notice latency spikes.

→ Configuration: MCP servers section — full schema.