Image Generation MCP

Multi-provider image generation MCP server built on FastMCP. Generate images from Claude Desktop, Claude Code, or any MCP client using OpenAI, Google Gemini, Stable Diffusion (SD WebUI), or a zero-cost placeholder provider.

Documentation | Config wizard | PyPI | Docker

Features

• Multi-provider: OpenAI (gpt-image-2, gpt-image-1.5, dall-e-3), Google Gemini (gemini-3.1-flash-image, gemini-3-pro-image, gemini-3.1-flash-lite-image), SD WebUI (Stable Diffusion / Forge / reForge), and a zero-cost placeholder for testing.
• Per-model style metadata: every model carries a style_profile (strengths, prompt grammar, lifecycle); list_providers includes a top-level warnings array for deprecated models. See Model Catalog.
• Keyword-based auto-selection: provider="auto" routes by prompt content (text/logo → OpenAI, photoreal/anime → SD WebUI, draft → placeholder).
• CDN-style image transforms: image://{id}/view?format=webp&width=512&crop_x=... resizes / re-encodes / crops on demand without re-generating.
• Hybrid background tasks: long-running SD generations run with task=True (poll for status); short OpenAI calls stream progress in the foreground.
• MCP Apps gallery + viewer: interactive UI surfaces (browse generated images, edit / crop / rotate) for clients that support app: resources.
• Production deployment: Docker (multi-arch), .deb/.rpm with hardened systemd, OIDC + bearer auth, persistent EventStore for HTTP session resumability.

What you can do with it

With this server mounted in an MCP client, you can ask:

• "Generate a coffee mug product photo on a worn oak table, 16:9, no text." Routes to gpt-image-1.5 for typography-aware photorealism.
• "Create three concept-art variations of a cyberpunk alley at dusk." Composes generate_image with provider="sd_webui" and a stylised checkpoint like dreamshaperXL.
• "Crop this image to a 1:1 square centred on the subject and resize to 512px." Uses image://{id}/view?width=512&height=512&crop_x=... resource transforms.
• "Show me my recent generations." Browses the gallery via the image://list resource and the MCP Apps gallery viewer.
• "Save this style as 'cyberpunk-night' so I can apply it to future requests." Uses the style library, whose markdown briefs the LLM interprets per-provider.
• "Replace the background of my last photo with a sunset sky." Uses transform_image with the gallery image_id as a reference (image-to-image via Gemini).

Installation

From PyPI

pip install image-generation-mcp

If you add optional extras via the PROJECT-EXTRAS-START / PROJECT-EXTRAS-END sentinels in pyproject.toml, document them below:

Example: pip install image-generation-mcp[all].

From source

git clone https://github.com/pvliesdonk/image-generation-mcp.git
cd image-generation-mcp
uv sync --all-extras --all-groups

Docker

docker pull ghcr.io/pvliesdonk/image-generation-mcp:latest

A compose.yml ships at the repo root as a starting point. Copy .env.example to .env, edit, and docker compose up -d.

To attach a remote Python debugger (development only; the protocol is unauthenticated), see Remote debugging.

Linux packages (.deb / .rpm)

Download .deb or .rpm packages from the GitHub Releases page. Both install a hardened systemd unit; env configuration is sourced from /etc/image-generation-mcp/env (copy from the shipped /etc/image-generation-mcp/env.example).

Claude Desktop (.mcpb bundle)

Download the .mcpb bundle from the GitHub Releases page and double-click to install, or run:

mcpb install image-generation-mcp-<version>.mcpb

Claude Desktop prompts for required env vars via a GUI wizard, with no manual JSON editing needed.

For manual Claude Desktop configuration and setup options, see Claude Desktop deployment.

Quick start

image-generation-mcp serve                                # stdio transport
image-generation-mcp serve --transport http --port 8000   # streamable HTTP

For library usage (embedding the domain logic without the MCP transport), import from the image_generation_mcp package directly. See the project's domain modules under src/image_generation_mcp/ for entry points.

Server info

The server registers a built-in get_server_info tool (via fastmcp_pvl_core.register_server_info_tool) so operators can confirm the deployed version with a single MCP call. The default response carries server_name, server_version, and core_version. Servers that talk to a remote upstream wire upstream version reporting inside the DOMAIN-UPSTREAM-START / DOMAIN-UPSTREAM-END sentinel in src/image_generation_mcp/server.py; see CLAUDE.md for the wiring pattern.

Configuration

Core environment variables shared across all fastmcp-pvl-core-based services:

Domain-specific variables go below under Domain configuration.

Authentication

Callers authenticate via a bearer token or OIDC (mutually exclusive). See the Authentication guide for setup, mapped multi-subject tokens, OIDC, and troubleshooting.

Post-scaffold checklist

After copier copy and gh repo create --push:

1. Fill in the DOMAIN blocks (every section marked with a DOMAIN sentinel comment) in this README and in CLAUDE.md.
2. Configure GitHub secrets (see below).
3. Install dev + docs tooling: uv sync --all-extras --all-groups.
4. Install pre-commit hooks: uv run pre-commit install.
5. Run the gate locally: uv run pytest -x -q && uv run ruff check --fix . && uv run ruff format . && uv run mypy src/ tests/.
6. Push the first commit. CI should be green.

GitHub secrets

CI workflows reference three repository secrets. Configure them via Settings → Secrets and variables → Actions or with gh secret set:

gh secret set RELEASE_TOKEN
gh secret set CODECOV_TOKEN
gh secret set CLAUDE_CODE_OAUTH_TOKEN

GITHUB_TOKEN is auto-provided; no action needed.

Local development

The PR gate (matches CI):

uv run pytest -x -q                                  # tests
uv run ruff check --fix . && uv run ruff format .    # lint + format
uv run mypy src/ tests/                              # type-check

Pre-commit runs a subset of the gate on each commit; see .pre-commit-config.yaml for details, or CLAUDE.md for the full Hard PR Acceptance Gates.

Troubleshooting

Moving a scaffolded project

uv sync creates .venv/bin/* scripts with absolute shebangs pointing at the venv Python. If you move the repo after scaffolding (mv /old/path /new/path), uv run pytest fails with ModuleNotFoundError: No module named 'fastmcp' because the stale shebang resolves to a different interpreter than the venv's site-packages.

Fix:

rm -rf .venv
uv sync --all-extras --all-groups

uv run python -m pytest also works as a one-shot workaround (bypasses the stale entry-script shim).

uv.lock refresh after copier update

When copier update introduces new dependencies (such as a new extra added to pyproject.toml.jinja), CI runs uv sync --frozen which fails against a stale lockfile. Run uv lock locally and commit the refreshed uv.lock alongside accepting the copier-update PR.

Links

• Documentation
• llms.txt
• FastMCP
• fastmcp-pvl-core

Domain configuration

All domain environment variables use the IMAGE_GENERATION_MCP_ prefix.

Core

Providers

Authentication

Cost control & performance

Reference image input (transform_image)

File Exchange (MCP downloads)

Server identity

Domain-config fields are composed inside src/image_generation_mcp/config.py between the CONFIG-FIELDS-START / CONFIG-FIELDS-END sentinels; env reads go through fastmcp_pvl_core.env(_ENV_PREFIX, "SUFFIX", default) so naming stays consistent.

For the full MCP tool / resource / prompt surface and per-provider setup notes, see the documentation site.

Key design decisions

• Multi-provider with capability discovery, not feature flags. Each provider's discover_capabilities() reports its actual supported aspect ratios / qualities / formats / negative-prompt support at startup; routing logic asks the capability surface, not a hard-coded enum. New providers slot in by satisfying the protocol, with no router edits needed. (See docs/decisions/0001-…, 0002-…, 0007-….)
• Per-model style_profile metadata, surfaced via list_providers. Closed-list providers (OpenAI, Gemini, placeholder) use exact-key lookup; SD WebUI uses a regex-ordered pattern table. Profiles include lifecycle flags (current / legacy / deprecated) and feed an auto-built top-level warnings array. (See docs/decisions/0009-….)
• Hybrid background tasks. Short calls (OpenAI ~5 s) stream progress in-line; long calls (SD WebUI 30-180 s) run as background tasks with check_generation_status polling; clients pick the mode via task=True. (See docs/decisions/0005-….)
• Image asset model: content-addressed registry + sidecar JSON metadata + on-demand transforms. Generated images keep their full-resolution original; image://{id}/view?format=webp&width=512&crop_x=… resources do format conversion / resize / crop on demand without re-generating. Transforms are cached. (See docs/decisions/0006-….)
• Style library. User-saved markdown briefs (with YAML frontmatter for tags / aspect ratio / quality) that the LLM interprets per-provider, not copy-pasted verbatim. Distinct from per-model style_profile: style library is the brief; style_profile describes the model. (See docs/decisions/0008-… and 0009-… for disambiguation.)
• Composes fastmcp_pvl_core.ServerConfig, never inherits. Domain config goes between CONFIG-FIELDS-START / CONFIG-FIELDS-END sentinels; env reads route through fastmcp_pvl_core.env(...) to keep prefix naming consistent.