MCP Server¶

The gateway can also run as an MCP server, so other MCP clients (Claude Desktop, Cursor, VS Code, another agent) can use GPT6-SOL as a tool.

Start it¶

mcp-for-copilot mcp

This speaks MCP over stdio. The process reads JSON-RPC from stdin and writes to stdout, so it must be launched by an MCP client rather than run interactively.

Exposed tools¶

Tool	Arguments	Returns
`gpt6_chat`	`prompt` (required), `model`, `system`	JSON object with `content`, `model`, `rounds`, `used_tools`, `plan`, `plan_progress`
`gpt6_models`	—	JSON object of model id → metadata
`gpt6_status`	—	JSON object of configuration booleans
`gpt6_tools`	—	JSON array of tools advertised by the connected MCP server
`gpt6_plan`	—	The working plan from the most recent `gpt6_chat` call
`gpt6_debug_marathon`	`tasks` (1-10 items)	Queue ordered debug work; returns `job_id`
`gpt6_debug_marathon_status`	`job_id`	Queue progress and model responses

Plan and verification fields¶

gpt6_chat always returns plan and plan_progress. Both are empty unless MCP_PLAN_MODE is enabled, in which case plan is a list of {"description", "status"} objects and plan_progress counts steps per status. gpt6_plan re-reads the plan from the last gpt6_chat call without re-running the model; it is read-only and never influences routing or approval.

gpt6_debug_marathon_status reports a verification field per task and per job. It is not_executed unless a stage was actually given execution output, in which case it is executed. The default is deliberately pessimistic: a model claiming success does not set executed. Each task also carries a plan list parsed from the model's MARATHON_PLAN line.

Each task accepts question plus optional title, priority, difficulty, and complexity (scores 1-5, default 3). Queue order is priority, difficulty, and complexity, descending. Per task, the default target paths are Luna-only for scores 1-2, Luna -> Sol for 3-4, and Luna -> Sol -> Astra for 5. An uncertain model result escalates regardless of the target tier. These defaults aim for an 80/15/5 task mix; they do not enforce quotas or guarantee a measured success rate. A completed result is model-self-reported and does not mean tests were run. Job state lives in the MCP process and is not durable across restarts. Status results include processed/completed counts, model-reported success rate, token usage, and estimated cost from the local model catalog.

Client configuration¶

Claude DesktopVS CodeCursor

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "gpt6-sol": {
      "command": "mcp-for-copilot",
      "args": ["mcp"],
      "env": {"OPENAI_API_KEY": "sk-..."}
    }
  }
}

Add to .vscode/mcp.json:

{
  "servers": {
    "gpt6-sol": {
      "type": "stdio",
      "command": "mcp-for-copilot",
      "args": ["mcp"],
      "env": {"OPENAI_API_KEY": "sk-..."}
    }
  }
}

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "gpt6-sol": {
      "command": "mcp-for-copilot",
      "args": ["mcp"],
      "env": {"OPENAI_API_KEY": "sk-..."}
    }
  }
}

Programmatic use¶

import asyncio
from mcp_for_copilot.gateway.server import build_server

server = build_server()
asyncio.run(server.run_stdio_async())

Verifying the server¶

The test suite includes an end-to-end test that spawns this server as a real subprocess and performs a full MCP handshake over stdio:

pytest tests/e2e -v

Notes¶

The server uses the mcp 2.x callback API (on_list_tools / on_call_tool passed to the Server constructor). The older decorator API (@server.list_tools()) was removed in mcp 2.0.
gpt6_chat requires a non-empty prompt; an empty value returns an error result rather than raising, so the client sees a normal tool error.
Tool results are returned as MCP TextContent blocks. gpt6_models, gpt6_status, and gpt6_tools return JSON-encoded strings.