GIMP MCP

Demo

Full demo (with audio): https://github.com/gimp-mcp/gimp-mcp/raw/main/docs/demo.mp4

AI agent using GIMP MCP to remove a background, edit a character's expression, and verify results — all through natural language via Claude

Overview

GIMP MCP bridges GIMP's professional image editing capabilities with AI assistants through the Model Context Protocol. It lets you edit images by describing what you want — and gives the AI a live visual feedback channel to verify each change before moving on.

What makes it different from other GIMP integrations:

• The AI can see the image at any point in the workflow without saving to disk (get_state_snapshot)
• Supports fully autonomous multi-step pipelines: open → edit → verify → refine → export
• 79 dedicated tool commands covering every major GIMP operation
• Fully compatible with GIMP 3.2.x (all breaking API changes resolved)

Key Features

What Can It Do?

Background Removal with Iterative Verification

The AI removes the background, takes a snapshot to inspect the result, detects remaining pixels, and loops until the image is clean:

"Remove the background from this image and keep looping until only the character remains"

Expression Editing

"Make the character smile — paint a smile arc with teeth over her mouth"

Complex Multi-Step Pipelines

"Open navi_portrait.png, remove the background, verify it's clean,
 then make her smile and export the final result as a PNG"

Color & Tone Work

"Boost the contrast, shift the hue 15 degrees warmer, then show me a before/after zoom of the face"

Text & Compositing

"Add a bold title at the top in white with a subtle drop shadow, then export for web"

Prerequisites

• GIMP 3.2+ — tested on GIMP 3.2.2 (Windows, macOS, Linux)
• uv — Python package manager (install uv); uvx downloads the MCP server and a suitable Python (3.11+) on first use
• MCP-compatible AI client — Claude Desktop, Claude Code, Gemini CLI, PydanticAI, etc.

Quick Start

The MCP server and the GIMP plugin ship together as the gimp-mcp-plugin package on PyPI. No clone needed.

1. Install the GIMP Plugin

Launch GIMP once (so its config folder exists), close it, then run:

uvx gimp-mcp-plugin@latest install-plugin

This copies the plugin into the newest GIMP 3.x config folder it finds (Linux standard, Snap and Flatpak locations, macOS, Windows) and makes it executable. Restart GIMP afterward.

• --dry-run shows where it would install without writing anything.
• --gimp-dir DIR installs into a specific GIMP config folder, e.g. --gimp-dir ~/.config/GIMP/3.2.

If you have more than one GIMP 3 installation (e.g. a system GIMP and a Flatpak), the newest folder found may not be the one your GIMP uses. Check the folder shown in GIMP's Edit → Preferences → Folders → Plug-ins and pass the folder that contains plug-ins with --gimp-dir.

Which folder does GIMP use? GIMP names its per-user folder after its major.minor version (3.0, 3.2, 3.4, …) and creates a fresh one on each minor upgrade, so the plug-ins folder moves when you upgrade GIMP. Re-run install-plugin after upgrading. To check the active path, open Edit → Preferences → Folders → Plug-ins in GIMP.

2. Start the MCP Server in GIMP

1. Open any image in GIMP
2. Go to Tools > MCP > Start MCP Server
3. Server starts on localhost:9877

Start automatically when GIMP launches

To skip step 2 on every launch, open Tools > MCP > MCP Server Settings..., tick Start MCP server when GIMP launches, and restart GIMP. Autostart is off by default.

The same dialog sets the port (default 9877). If you change it, set GIMP_MCP_PORT in your MCP client config so the MCP server connects to the same port:

"gimp": {
  "command": "uvx",
  "args": ["gimp-mcp-plugin@latest"],
  "env": { "GIMP_MCP_PORT": "9988" }
}

Settings are saved to gimp-mcp.json in GIMP's per-user config folder (e.g. ~/.config/GIMP/3.2/gimp-mcp.json). Port changes apply the next time the server starts.

Timeout for long commands

The MCP server waits up to 300 seconds for GIMP to finish each command. For heavier work (large or 16-bit images, slow filters), set GIMP_MCP_TIMEOUT (in seconds) in your MCP client config, the same way as GIMP_MCP_PORT above, e.g. "GIMP_MCP_TIMEOUT": "900". check_server reports the timeout in effect. When a command times out, GIMP may still be running it, so check the image before retrying. Your MCP client may have its own tool timeout (in Claude Code, MCP_TOOL_TIMEOUT, in milliseconds); raise it too if it's shorter.

3. Configure Your MCP Client

gimp-mcp-plugin@latest makes uvx check PyPI for a new release each time your client starts the server.

Claude Desktop

~/.config/Claude/claude_desktop_config.json (Linux/macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "gimp": {
      "command": "uvx",
      "args": ["gimp-mcp-plugin@latest"]
    }
  }
}

Claude Code

claude mcp add gimp-mcp -- uvx gimp-mcp-plugin@latest

Gemini CLI

~/.config/gemini/.gemini_config.json:

{
  "mcpServers": {
    "gimp": {
      "command": "uvx",
      "args": ["gimp-mcp-plugin@latest"]
    }
  }
}

PydanticAI

from pydantic_ai import Agent
from pydantic_ai.mcp import MCPServerStdio

server = MCPServerStdio('uvx', args=['gimp-mcp-plugin@latest'])
agent = Agent('openai:gpt-4o', mcp_servers=[server])

Cancelling long commands

GIMP runs one command at a time, in the order they're sent. While it works, the MCP server stays responsive, so your client can cancel a call. The cancel returns at once: a command that's still waiting its turn never runs, but one GIMP has already started is finished in the background. Check the image before retrying a cancelled edit, or it may be applied twice.

Updating

The plugin and the MCP server must be the same version. check_server and get_gimp_info report a mismatch and print the command that fixes it. To update both:

1. Restart your MCP client. With @latest, it picks up the newest server.
2. Run uvx gimp-mcp-plugin@latest install-plugin and restart GIMP.

Running from a Git Checkout (contributors)

To work on the code, run the server and install the plugin from a clone instead:

git clone https://github.com/gimp-mcp/gimp-mcp.git
cd gimp-mcp
uv sync
uv run gimp_mcp_server.py install-plugin    # installs this checkout's plugin

Then point your MCP client at the checkout, e.g. for Claude Desktop:

"gimp": {
  "command": "uv",
  "args": ["run", "--directory", "/full/path/to/gimp-mcp", "gimp_mcp_server.py"]
}

Claude Code picks up the repo's .mcp.json automatically when started in the checkout. After git pull, re-run install-plugin if the plugin changed.

Available MCP Tools

👁️ Visual Feedback

get_state_snapshot(image_index, max_size, region, label)

Returns a live PNG of the current image state — the AI's primary feedback mechanism. Call this between any edits to verify the result without saving to disk.

# Full image snapshot
snapshot = get_state_snapshot(max_size=512)

# Zoom into a face region for detail inspection
snapshot = get_state_snapshot(
    region={"x": 140, "y": 80, "width": 240, "height": 300},
    max_size=512,
    label="face-check"
)

This enables iterative agentic workflows: edit → snapshot → assess → refine → repeat.

get_image_bitmap(image_index, max_width, max_height, region)

Lower-level bitmap fetch with region extraction and scaling. Returns base64-encoded PNG.

🎨 Adjustments

🔄 Transforms

✂️ Selections

🗂️ Layers

🖌️ Drawing & Fill

🔤 Text

✨ Filters & Effects

📁 File Operations

📦 Batch & Multi-Format Export

↩️ History

🔍 Info & Context

🔌 Connection & Raw API

AI Agent Feedback Loop

The get_state_snapshot tool enables a pattern where the AI loops until a goal is visually confirmed:

┌─────────────┐
│  Apply edit │
└──────┬──────┘
       │
       ▼
┌─────────────────┐
│ get_state_      │  ← AI sees live PNG, no disk save needed
│ snapshot()      │
└──────┬──────────┘
       │
       ▼
┌─────────────────┐     ┌──────────────────┐
│ Goal achieved?  │─ No─▶ Adjust & retry   │
└──────┬──────────┘     └──────────────────┘
       │ Yes
       ▼
┌─────────────┐
│   Export    │
└─────────────┘

Example: Iterative Background Removal

See bg_remove_iterative.py for a complete example. The AI:

1. Removes the background using edge-seeded contiguous select
2. Takes a snapshot to check the result
3. Scans for remaining background-colored pixels
4. Runs targeted removal passes with progressively finer grids (25px → 1px)
5. Runs a final despeckle pass for isolated pixels
6. Loops until no background pixels remain

Example Scripts

Run the test suite to verify your setup:

python run_tests.py
# Expected: 59/59 PASSED

Technical Architecture

Plugin ↔ Server Communication

AI Client (Claude, etc.)
      │  MCP (stdio)
      ▼
gimp_mcp_server.py          ← MCP tool definitions
      │  TCP JSON  :9877
      ▼
gimp-mcp-plugin.py          ← Runs inside GIMP process
      │  PyGObject
      ▼
GIMP 3.2 (gi.repository.Gimp)

• MCP server translates tool calls into JSON commands sent to the plugin over TCP
• Plugin executes operations directly in the GIMP process via PyGObject
• Two message formats: {"type": "...", "params": {...}} for named tools, {"cmds": ["python..."]} for arbitrary exec

GIMP 3.2 Compatibility Notes

GIMP 3.x introduced breaking API changes from GIMP 2.x. Key fixes included in this release:

Troubleshooting

"Could not connect to GIMP"

• GIMP must be running with an image open
• Start the MCP server: Tools > MCP > Start MCP Server, or enable autostart in Tools > MCP > MCP Server Settings...
• If you changed the port in GIMP, set GIMP_MCP_PORT to the same value in your MCP client config
• Check port 9877 (or your custom port) is not blocked by firewall

"GIMP did not reply within 300 s"

• The command took longer than the MCP server waits. Raise GIMP_MCP_TIMEOUT (seconds) in your MCP client config; see Timeout for long commands
• GIMP may still be running the command: check the image before retrying

Plugin Not Visible in GIMP

• Look under Tools > MCP (the plugin adds an MCP submenu, not a top-level Tools entry)
• Confirm the plugin file is in the correct directory: uvx gimp-mcp-plugin@latest install-plugin --dry-run shows where it goes
• Upgraded GIMP recently? A minor upgrade (e.g. 3.0 → 3.2) moves the per-user config folder to a new version directory; re-run uvx gimp-mcp-plugin@latest install-plugin (use --dry-run to see which folder it picks). Verify the active path via Edit > Preferences > Folders > Plug-ins.
• On Linux/macOS: ensure the file has execute permission (chmod +x)
• Restart GIMP after installation
• Check Filters > Script-Fu > Console for error messages

Tests Failing

Run the test suite and check the failure list:

python run_tests.py

Each failure includes the tool name and error — most issues on GIMP 3.2 are covered by the fixes above.

Example Output

"Draw me a face and a sheep" — generated entirely through natural language via GIMP MCP

Future Enhancements

• 📚 Recipe Collection: Reusable workflow templates (portrait cleanup, product photo, etc.)
• 🚀 Dynamic Discovery: Auto-generate MCP tools from GIMP's full PDB procedure database
• 🔒 Security: Sandboxed execution for untrusted command inputs
• ⚡ Performance: Optimized bitmap transfer for large images
• 🌐 Remote Access: Network-accessible GIMP instances

Contributing

Contributions are welcome — bug fixes, new tools, documentation, or example scripts. Open a PR or issue on GitHub.

Development Setup

Install dev dependencies and activate the pre-commit hook so ruff runs on every commit:

uv sync
uv run pre-commit install

After this, ruff checks staged files on each git commit (with --fix applied automatically). The same check runs in CI, so the hook is just a fast local safety net.

To bump the pinned hook versions later:

uv run pre-commit autoupdate

License

GPL-3.0-or-later. See LICENSE.