Copilot Image Generation MCP

An MCP server that connects GitHub Copilot CLI to Microsoft Copilot's image generation backend.

Generate images from text prompts and iteratively refine them — all from the terminal. Images are saved locally and organized by session.

Requirements

• Python 3.10+
• GitHub Copilot CLI
• Microsoft 365 Copilot license with image generation enabled
• A Chromium-based browser (Edge, Chrome, or Chromium) for first-time sign-in

Quick Start

1. Install and register

   pip install copilot-image-gen-mcp
   copilot plugin install msartem/copilot_image_gen_mcp

1. Launch Copilot CLI

   copilot

1. Sign in (first time only)

Copilot will automatically call sign_in when needed. An Edge browser window opens — sign in with your Microsoft 365 account. The auth code is captured automatically via Playwright (no manual copy-paste).

1. Try it out

   Generate an image of an elephant in Times Square
   Make the elephant golden
   Change the background to a sunset over mountains

After first sign-in, auth is silent via cached refresh tokens (~90 day lifetime, auto-renewing). No browser window on subsequent uses.

How It Works

1. You ask Copilot CLI to generate or modify an image
2. Copilot routes it to the generate_image or refine_image tool
3. The MCP server opens a WebSocket to Microsoft Copilot's backend
4. The backend generates the image (DALL-E) and returns it as base64 PNG
5. The image is saved to ~/.copilot-images/ and the file path is returned

Multi-turn refinement works automatically — the server maintains a conversation so each refine_image call builds on the previous image.

You → Copilot CLI → image gen MCP → M365 Copilot (Sydney) → DALL-E → PNG saved locally

Tools

Orientation Options

• landscape (default)
• portrait
• square

Image Storage

Images are saved to ~/.copilot-images/ organized by session. Each session groups an initial image with its refinements:

~/.copilot-images/
├── 20250115_143022_elephant_in_times_square/
│   ├── session.json                              ← metadata (prompts, timestamps)
│   ├── 001_elephant_in_times_square.png           ← initial image
│   ├── 002_make_the_elephant_golden.png           ← first refinement
│   └── 003_change_the_background_to_a_sunset.png  ← second refinement
└── 20250116_091200_sunset_over_mountains/
    ├── session.json
    └── 001_sunset_over_mountains.png

Session directories are created lazily on first image save — no side effects until you actually generate something.

Authentication

Authentication uses Playwright to automate browser sign-in via Microsoft Edge. This works identically on macOS and Windows — no platform-specific code.

First sign-in

1. Copilot calls sign_in (or you trigger it manually)
2. An Edge window opens to the Microsoft sign-in page
3. Sign in with your M365 account (SSO may auto-complete this)
4. The auth code is captured automatically — the browser closes
5. Tokens are cached locally

Subsequent runs

Cached refresh tokens are used — no browser window, no interaction needed.

Browser selection

Playwright uses Microsoft Edge by default. To use a different Chromium-based browser:

export COPILOT_BROWSER=chrome    # or: msedge (default), chromium

Note: chromium requires playwright install chromium. Edge and Chrome use your installed browser directly — no extra install needed.

Manual auth

python auth.py          # Interactive sign-in
python auth.py logout   # Clear cached tokens

Token storage

Environment Variables

Technical Details

See TECHNICAL.md for details on the SignalR WebSocket protocol, authentication flow, image delivery format, and multi-turn refinement architecture.

Disclaimer

This is an independent community project — not affiliated with or supported by Microsoft.

It works by communicating with the same undocumented web APIs that power the M365 Copilot web/desktop app image generation. These APIs may change or break without notice. Use at your own risk.