🛰️ MCP for Copilot

Give Copilot hands.

A production-ready Model Context Protocol gateway. Use it as a Python library or deploy it as an OpenAI-compatible FastAPI service.

Quick Start · Installation · Examples · Architecture · Docs · Contributing

🎯 What is this?

Copilot is a strong reasoning assistant, but on its own it cannot read your files, query your database, or run your tools. MCP for Copilot connects it to any MCP server and puts three safety gates in front of every tool call: an allowlist, an approval layer, and argument validation.

It speaks two protocols at once — MCP on the tool side, OpenAI on the client side — so any OpenAI SDK client can use it as a drop-in replacement, and any MCP server can be its tool provider.

Related MCP server: LangGraph FastAPI MCP Server

✨ Features

• 🔌 Three MCP transports — stdio, http (streamable), and sse

• 🧠 GPT-6 family aware — Sol, Luna, Astra, and GPT-5.6 Sol/Luna with correct context windows, output caps, and pricing metadata

• 🛡️ Three-gate tool safety — allowlist → approval → dispatch, and it fails closed at every gate

• ✅ Argument validation — JSON-schema subset validation before dispatch, so a malformed call never reaches your server

• 🚦 Risky-tool detection — 40+ mutating verbs (write, delete, exec, …) are flagged automatically; unknown tools are treated as risky

• 🔐 Secret redaction — 11 regex rules scrub keys, tokens, and JWTs from every log line and error message

• 🧯 Prompt-injection guard — tool output is wrapped and labelled as untrusted data before it re-enters the model context

• 🌊 Streaming — OpenAI-compatible SSE with the exact chunk format clients expect

• 🐳 Multi-stage Docker — one image, both modes, non-root user, healthcheck

• 🧩 Zero-config library mode — pip install and go; no server required

• 🧪 Typed and tested — full type hints, mypy --strict, 3-layer test suite

🚀 Quick Start

Option A — Use as a library

pip install "mcp-for-copilot[library]"

import asyncio
from mcp_for_copilot import Gateway

async def main():
    async with Gateway() as gateway:
        result = await gateway.chat([
            {"role": "user", "content": "List the files in the current directory."}
        ])
        print(result.content)
        print("tools used:", result.used_tools)

asyncio.run(main())

Option B — Deploy as a server

pip install "mcp-for-copilot[server]"
export LLM_API_KEY=sk-...
mcp-for-copilot serve

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-6-sol", "messages": [{"role": "user", "content": "Hello!"}]}'

That is the whole setup. Any OpenAI SDK client works against it unchanged.

📦 Installation

pip

# Library only — talk to an MCP server from your own code
pip install "mcp-for-copilot[library]"

# Server — FastAPI gateway + stdio MCP server
pip install "mcp-for-copilot[server]"

# Everything
pip install "mcp-for-copilot[all]"

Docker

docker run --rm -p 8000:8000 \
  -e LLM_API_KEY=sk-... \
  ghcr.io/rwamyth-blip/mcp-for-copilot:latest

From source

git clone https://github.com/rwamyth-blip/mcp-for-copilot.git
cd mcp-for-copilot
pip install -e ".[all,dev]"
cp .env.example .env   # then edit .env

💡 Usage Examples

1. Library — chat with tools, with an audit trail

import asyncio
from mcp_for_copilot import Gateway

async def main():
    async with Gateway() as gateway:
        result = await gateway.chat(
            [{"role": "user", "content": "How many documents are in the users collection?"}],
            system="You are a helpful data assistant.",
        )

        print(result.content)
        for call in result.invocations:
            status = "ok" if call["executed"] and not call["is_error"] else "blocked"
            print(f"  [{status}] {call['tool']} — {call['reason']}")

asyncio.run(main())

2. Server — OpenAI SDK, unchanged

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed-locally")

response = client.chat.completions.create(
    model="sol",  # alias for gpt-6-sol
    messages=[{"role": "user", "content": "Summarise README.md"}],
)
print(response.choices[0].message.content)

3. Custom approval policy — approve only what you trust

import asyncio
from mcp_for_copilot import Gateway, static_approver

async def main():
    approver = static_approver(allow=["write_file"], deny=["delete_file"])
    async with Gateway(approver=approver) as gateway:
        result = await gateway.chat([
            {"role": "user", "content": "Write 'hello' to notes.txt"}
        ])
        print(result.content)

asyncio.run(main())

4. CLI — one-shot prompt

mcp-for-copilot chat "What models do you support?" --json
mcp-for-copilot models
mcp-for-copilot status

🏗️ Architecture

flowchart LR
    subgraph Client["Client side"]
        A["OpenAI SDK<br/>/ any HTTP client"]
        B["Your Python code"]
    end

    subgraph Gateway["mcp-for-copilot"]
        C["FastAPI app<br/>/v1/chat/completions"]
        D["Gateway facade"]
        E["Orchestrator<br/>model → tool → model"]
        F["ToolRouter<br/>allowlist + schema"]
        G["ApprovalLayer<br/>fail closed"]
        H["LLMProvider<br/>OpenAI adapter"]
        I["MCPClient<br/>stdio / http / sse"]
    end

    subgraph Tools["Tool side"]
        J["MCP server"]
        K["Files, DB, APIs"]
    end

    L["GPT-6 Sol"]

    A --> C --> D --> E
    B --> D
    E --> H --> L
    E --> F --> G --> I --> J --> K
    J -. "tool schemas" .-> F

The three gates

Every tool call the model requests must pass all three, in order:

A denied tool never aborts the turn. The model is told why it was denied and can explain the outcome or try a different approach.

Why reasoning_effort is forced to none with tools

The GPT-6 family rejects function tools combined with reasoning_effort on /v1/chat/completions:

Function tools with reasoning_effort are not supported for gpt-6-luna in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.

The field must be present and set to "none" — omitting it is also rejected. Since tool calling is the entire point of this package, the provider adapter sends "none" whenever tools are present, and forwards your configured effort unchanged when they are not.

⚙️ Configuration

All configuration is environment-driven. See .env.example for the annotated list.

Supported models

Aliases: gpt-6, gpt6, sol → gpt-6-sol · luna → gpt-6-luna · astra → gpt-6-astra · gpt-5.6 → gpt-5.6-sol

🐳 Docker

# Gateway mode (default)
docker compose up gateway

# stdio MCP server mode
docker compose up mcp

# Both
docker compose up

The image is multi-stage: dependencies are built in a builder stage, and the runtime stage runs as a non-root user with a HEALTHCHECK on /health.

🧪 Testing

make test          # pytest with coverage
make lint          # ruff check + format check
make typecheck     # mypy --strict
make check         # all of the above

Tests are split into three layers:

No test touches the network: httpx transports and the MCP client are injected.

📚 Documentation

Full documentation lives at rwamyth-blip.github.io/mcp-for-copilot.

• Getting started

• Library guide

• Server guide

• Security model

• API reference

🤝 Contributing

Contributions are welcome. Please read CONTRIBUTING.md and CODE_OF_CONDUCT.md first.

git clone https://github.com/rwamyth-blip/mcp-for-copilot.git
cd mcp-for-copilot
pip install -e ".[all,dev]"
pre-commit install
make check

Good first issues are labelled good first issue.

🔐 Security

Please do not open a public issue for a vulnerability. See SECURITY.md for the private disclosure process.

This package never logs a credential: every log line and error message passes through mcp_for_copilot.logging_utils.redact().

📄 License

MIT © VihokAI

⭐ Star History

⬆ back to top