MCP server

trackfox mcp start runs the CLI as a local Model Context Protocol server over stdio. Unlike the AI agent skill, which teaches an agent to run CLI commands, an MCP server exposes TrackFox as structured tools an MCP client calls directly — so it also works in clients that can't run arbitrary shell commands, like Claude Desktop.

Add it to Claude Desktop or Claude Code

Point your client's MCP config at the CLI with credentials in the environment:

{
  "mcpServers": {
    "trackfox": {
      "command": "trackfox",
      "args": ["mcp", "start"],
      "env": {
        "TRACKFOX_WEBSITE_ID": "site_123",
        "TRACKFOX_API_KEY": "tf_secret"
      }
    }
  }
}

You can also omit the env block and pass a saved profile name as an argument ("args": ["mcp", "start", "my-profile"]) if you've already run trackfox config add. See Profiles for managing multiple websites.

What it exposes

The server registers one tool per CLI capability — the same analytics reads (trackfox_metadata, trackfox_chart, trackfox_realtime, trackfox_pages, trackfox_referrers, trackfox_locations, trackfox_platforms, trackfox_custom_events, trackfox_electron_versions) plus the management actions covered in Manage automations, funnels & segments (trackfox_get_automation, trackfox_update_discord_automation, trackfox_update_telegram_automation, trackfox_list_funnels, trackfox_create_funnel, trackfox_delete_funnel, trackfox_list_segments, trackfox_create_segment, trackfox_delete_segment, trackfox_attribute_payment). Run trackfox mcp start and call tools/list from any MCP client to see the exact schema for each.

Good to know

• The server only talks MCP over stdio — nothing is written to stdout except the protocol itself. Status messages go to stderr, so they won't corrupt the connection if you run it manually to check it starts.
• It authenticates exactly like the CLI: a saved profile, or TRACKFOX_WEBSITE_ID/TRACKFOX_API_KEY in the environment. There's no separate MCP-specific credential.
• This is a local server — it runs on your machine, started by your MCP client. It isn't a hosted connector you can add from claude.ai without running the CLI yourself.
• The management tools write to your site. A well-behaved MCP client will ask before calling one that changes or deletes something, but the API key itself has full access to those actions — scope who has that key accordingly.