An AI assistant that can
actually see your game.

Launch FRED-80 with a token

$ export FRED80_MCP_TOKEN=some-long-random-string
$ ./fred-80
> :: Windows (cmd)
> set FRED80_MCP_TOKEN=some-long-random-string
> fred-80.exe

The MCP server listens on 127.0.0.1:8090 and only accepts connections that present this token. macOS and Linux since the start; Windows support (Winsock2) landed in 0.8.200, verified compiling and linking clean through the real mingw-w64 cross-toolchain -- not yet confirmed on actual Windows hardware, so treat it as recently-added rather than battle-tested. Amiga/Apollo hardware transport is still a separate, later, opt-in phase and remains stubbed for now. On Windows specifically: the GUI build has no console, so the auto-generated token (if you didn't set one yourself) is written to fred80.log next to the exe instead of the terminal -- check there if you launched without setting FRED80_MCP_TOKEN first. Add --headless to run with no window at all (everything, including screenshot, already reads the internal framebuffer rather than the displayed window, so nothing else changes) -- useful for pure AI-automation with no one watching.

Point your AI host at fred-80 --mcp-bridge

{
  "mcpServers": {
    "fred80": {
      "command": "/path/to/fred-80",
      "args": ["--mcp-bridge"],
      "env": { "FRED80_MCP_TOKEN": "some-long-random-string" }
    }
  }
}

--mcp-bridge connects to an already-running fred-80 instance (the one you launched in step 1 -- it never starts a second engine of its own) and relays stdin/stdout to its MCP socket, exactly like a real MCP host expects. This is the same binary either way -- there's nothing extra to download or install, on any platform. Prefer a Python-based setup instead? tools/mcp_bridge.py in the repo still works identically, speaking the same wire protocol.

Just ask

Open a cart and describe what you want. The assistant reaches for screenshot, sprite_get, or get_error_log the same way it reaches for any other tool — no special prompting required.