🎨 Image Gen MCP Server

"Fine. I'll do it myself." — Thanos (and also me, after trying five different MCP servers that couldn't mix-and-match image models)
I wanted a single, simple MCP server that lets agents generate and edit images across OpenAI, Google (Gemini/Imagen), Azure, Vertex, and OpenRouter—without yak‑shaving. So… here it is.

A multi‑provider Model Context Protocol (MCP) server for image generation and editing with a unified, type‑safe API. It returns MCP ImageContent blocks plus compact structured JSON so your client can route, log, or inspect results cleanly.

🗺️ Table of Contents

• Why this exists

• Features

• Quick start (users)

• Quick start (developers)

• Configure mcp.json

• Tools API

  • generate_image

  • edit_image

  • get_model_capabilities

• Providers & Models

• Python client example

• Environment Variables

• Running via FastMCP CLI

• Troubleshooting & FAQ

• Contributing & Releases

• License

🧠 Why this exists

Because I couldn’t find an MCP server that spoke multiple image providers with one sane schema. Some only generated, some only edited, some required summoning three different CLIs at midnight.
This one prioritizes:

• One schema across providers (AR & diffusion)

• Minimal setup (uvx or pip, drop a mcp.json, done)

• Type‑safe I/O with clear error shapes

• Discoverability: ask the server what models are live via get_model_capabilities

✨ Features

• Unified tools: generate_image, edit_image, get_model_capabilities

• Providers: OpenAI, Azure OpenAI, Google Gemini, Vertex AI (Imagen & Gemini), OpenRouter

• Output: MCP ImageContent blocks + small JSON metadata

• Quality/size/orientation normalization

• Masking support where engines allow it

• Fail‑soft errors with stable shape: { code, message, details? }

🚀 Quick start (users)

Install and use as a published package.

# With uv (recommended)
uv add image-gen-mcp

# Or with pip
pip install image-gen-mcp

Then configure your MCP client.

Configure mcp.json

Use uvx to run in an isolated env with correct deps:

{
  "mcpServers": {
    "image-gen-mcp": {
      "command": "uvx",
      "args": ["--from", "image-gen-mcp", "image-gen-mcp"],
      "env": {
        "OPENAI_API_KEY": "your-key-here"
      }
    }
  }
}

First call

{
  "tool": "generate_image",
  "params": {
    "prompt": "A vibrant painting of a fox in a sunflower field",
    "provider": "openai",
    "model": "gpt-image-1"
  }
}

🧑‍💻 Quick start (developers)

Run from source for local development or contributions.

Prereqs

• Python 3.12+

• uv (recommended)

Install deps

uv sync --all-extras --dev

Environment

cp .env.example .env
# Add your keys

Run the server

# stdio (direct)
python -m image_gen_mcp.main

# via FastMCP CLI
fastmcp run image_gen_mcp/main.py:app

Local VS Code mcp.json for testing

If you use a VS Code extension or local tooling that reads .vscode/mcp.json, here's a safe example to run the local server (do NOT commit secrets):

{
  "servers": {
    "image-gen-mcp": {
      "command": "python",
      "args": ["-m", "image_gen_mcp.main"],
      "env": {
        "# NOTE": "Replace with your local keys for testing; do not commit.",
        "OPENROUTER_API_KEY": "__REPLACE_WITH_YOUR_KEY__"
      }
    }
  },
  "inputs": []
}

Use this to run the server from your workspace instead of installing the package from PyPI. For CI or shared repos, store secrets in the environment or a secret manager and avoid checking them into git.

Dev tasks

uv run pytest -v
uv run ruff check .
uv run black --check .
uv run pyright

🧰 Tools API

All tools take named parameters. Outputs include structured JSON (for metadata/errors) and MCP ImageContent blocks (for actual images).

generate_image

Create one or more images from a text prompt.

Example

{
  "prompt": "A vibrant painting of a fox in a sunflower field",
  "provider": "openai",
  "model": "gpt-image-1",
  "n": 2,
  "size": "M",
  "orientation": "landscape"
}

Parameters

edit_image

Edit an image with a prompt and optional mask.

Example

{
  "prompt": "Remove the background and make the subject wear a red scarf",
  "provider": "openai",
  "model": "gpt-image-1",
  "images": ["data:image/png;base64,..."],
  "mask": null
}

Parameters

get_model_capabilities

Discover which providers/models are actually enabled based on your environment.

Example

{ "provider": "openai" }

Call with no params to list all enabled providers/models.

Output: a CapabilitiesResponse with providers, models, and features.

🧭 Providers & Models

Routing is handled by a ModelFactory that maps model → engine. A compact, curated list keeps things understandable.

Model Matrix

Provider Model Support

🐍 Python client example

import asyncio
from fastmcp import Client


async def main():
    # Assumes the server is running via: python -m image_gen_mcp.main
    async with Client("image_gen_mcp/main.py") as client:
        # 1) Capabilities
        caps = await client.call_tool("get_model_capabilities")
        print("Capabilities:", caps.structured_content or caps.text)

        # 2) Generate
        gen_result = await client.call_tool(
            "generate_image",
            {
                "prompt": "a watercolor fox in a forest, soft light",
                "provider": "openai",
                "model": "gpt-image-1",
            },
        )
        print("Generate Result:", gen_result.structured_content)
        print("Image blocks:", len(gen_result.content))


asyncio.run(main())

🔐 Environment variables

Set only what you need:

🏃 Running via FastMCP CLI

Supports multiple transports:

• stdio: fastmcp run image_gen_mcp/main.py:app

• SSE (HTTP): fastmcp run image_gen_mcp/main.py:app --transport sse --host 127.0.0.1 --port 8000

• HTTP: fastmcp run image_gen_mcp/main.py:app --transport http --host 127.0.0.1 --port 8000 --path /mcp

Design notes

• Schema: public contract in image_gen_mcp/schema.py (Pydantic).

• Engines: modular adapters in image_gen_mcp/engines/, selected by ModelFactory.

• Capabilities: discovered dynamically via image_gen_mcp/settings.py.

• Errors: stable JSON error { code, message, details? }.

⚠️ Testing remarks

I tested this project locally using the openrouter-backed model only. I could not access Gemini or OpenAI from my location (Hong Kong) due to regional restrictions — thanks, US government — so I couldn't fully exercise those providers.

Because of that limitation, the gemini/vertex and openai (including Azure) adapters may contain bugs or untested edge cases. If you use those providers and find issues, please open an issue or, even better, submit a pull request with a fix — contributions are welcome.

Suggested info to include when filing an issue:

• Your provider and model (e.g., openai:gpt-image-1, vertex:imagen-4.0-generate-001)

• Full stderr/server logs showing the error

• Minimal reproduction steps or a short test script

Thanks — and PRs welcome!

🤝 Contributing & Releases

PRs welcome! Please run tests and linters locally.

Release process (GitHub Actions)

1. Automated (recommended)

   • Actions → Manual Release

   • Pick version bump: patch / minor / major

   • The workflow tags, builds the changelog, and publishes to PyPI

2. Manual

   • git tag vX.Y.Z

   • git push origin vX.Y.Z

   • Create a GitHub Release from the tag

📄 License

Apache-2.0 — see LICENSE.