AI coding agents are stateless and repo-scoped. Your fleet isn't. fleet-mcp is an opinionated MCP server that maintains a living Knowledge Graph of your multi-app SaaS architecture — modelling the relationships between services, the contracts at their boundaries, the shared concerns that cross them, and the things that silently break other apps when touched naively. Wire it once; every agent session across every codebase in your fleet inherits full situational awareness.

Wiring to Claude Desktop

stdio (local, recommended for Phase 1)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or equivalent:

{
  "mcpServers": {
    "fleet": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/fleet-mcp"
    }
  }
}

Restart Claude Desktop after editing. The five fleet_* tools will be available in every session.

HTTP (optional, for team sharing)

mcp run src/server.py --transport streamable-http --port 8096

Point Claude Desktop at http://localhost:8096/mcp.

Development Workflow

Changes to this project follow the Harnessable agent development pipeline: /engineer → /coder → /qa. See docs/harness/QUICKSTART.md for the full quickstart.