gimp-mcp
Enables scripted image editing with GIMP 3, including cropping, resizing, aspect-ratio fitting, dimension-spec validation, batch processing, and live control of a running GIMP instance.
README
gimp-mcp
An MCP server that drives GIMP 3 for scripted image editing: crop, resize, aspect-ratio fitting, light colour touch-up, dimension-spec validation, and batch processing across a folder.
Built and verified on Windows with GIMP 3.2.4, using GIMP 3's
GObject Introspection Python API (gi.repository.Gimp) rather than the old
2.x Script-Fu interface.
What it is for
Any workflow where images need the same deterministic treatment applied repeatedly and you would rather describe it than click through it:
- crop a photo to a target aspect ratio, or to the largest centred square
- resize a folder of images so the longest edge is at most 2000px
- check whether images meet a size/orientation requirement before publishing
- apply one crop-and-resize pipeline across a whole shoot in one pass
The one thing that will bite you: EXIF orientation
Photos from phones and many cameras are frequently stored landscape with an EXIF orientation tag telling viewers to rotate them. A photo everyone sees as 3000x4000 portrait may be stored as 4000x3000.
GIMP's non-interactive loader does not apply that tag. A naive "crop to square, centered" therefore crops the wrong axis and produces a sideways image — while still reporting plausible-looking dimensions, so nothing looks obviously broken until you open the output.
Every load in this project goes through load_image(), which calls
Gimp.Image.policy_rotate() first, so all geometry — and every dimension this
server reports — is in displayed orientation, i.e. what a viewer actually
sees. This is covered by a test.
Architecture
Two execution backends, one shared operation runtime:
┌───────────────────────────────┐
MCP client ──────►│ gimp_mcp/server.py (stdio) │
└───────────┬───────────────────┘
│
┌─────────────────┴──────────────────┐
▼ ▼
HeadlessBackend BridgeBackend
spawns gimp-console-3.exe TCP 127.0.0.1:50472
(no running GIMP needed) (into a running GIMP)
│ │
▼ ▼
bootstrap.py plug-ins/gimp-mcp-bridge/
│ │
└──────────────┬─────────────────────┘
▼
gimp_mcp/gimp_runtime.py
THE single source of truth for every
image operation. Both paths share it,
so batch and live cannot drift apart.
install_plugin.py writes a runtime_path.txt pointer next to the installed
plug-in rather than copying gimp_runtime.py, so exactly one copy of the
operation code exists on disk.
Backend choice. headless is the default and is what all batch and
deterministic work uses — it needs no open GIMP and is the reliable path.
bridge is for live work on a document you already have open. Both are
verified to produce pixel-identical output.
Why TCP and not D-Bus
Existing live-GIMP-control projects use D-Bus, which does not exist on
Windows. A loopback TCP socket achieves the same thing and is
cross-platform. It binds 127.0.0.1 only and is never exposed to the
network.
Install
Requires GIMP 3.x (developed against 3.2.4) and the mcp Python package.
Note on the
mcpdependency. This targets themcp1.x SDK and is pinned tomcp>=1.0,<2. Version 2.0 removedmcp.server.fastmcpand renamedFastMCPtoMCPServer; porting to it is not done yet, and an unpinned install picks up 2.x and fails at import.
pip install -r requirements.txt
python install_plugin.py # install the bridge plug-in (optional)
python install_plugin.py --list # show detected GIMP config dirs
The bridge plug-in is only needed for the live control tools. The batch and single-image tools work without installing anything into GIMP.
Plug-in location
install_plugin.py discovers whatever GIMP 3.x config directories actually
exist rather than hardcoding a version. On Windows that is:
%APPDATA%\GIMP\3.2\plug-ins\gimp-mcp-bridge\gimp-mcp-bridge.py
Note it is the versioned directory (3.2 for GIMP 3.2, not 3.0), and
GIMP 3 requires each plug-in to sit in a folder whose name matches the .py
file. On Linux and macOS the installer looks in ~/.config/GIMP/3.x/ and
~/Library/Application Support/GIMP/3.x/ respectively.
Register the MCP server
Installing the package provides a gimp-mcp console script, which is the
tidiest thing to register because it does not depend on a working directory:
python -m venv .venv
.venv/Scripts/python -m pip install -e . # .venv/bin/python on Unix
{
"mcpServers": {
"gimp": {
"type": "stdio",
"command": "/path/to/gimp-mcp/.venv/Scripts/gimp-mcp.exe",
"args": []
}
}
}
With Claude Code, the equivalent one-liner is:
claude mcp add gimp --scope user -- /path/to/gimp-mcp/.venv/Scripts/gimp-mcp.exe
Running the module directly works too, if mcp is importable in that
interpreter:
{
"mcpServers": {
"gimp": {
"command": "python",
"args": ["-m", "gimp_mcp"],
"cwd": "/path/to/gimp-mcp"
}
}
}
Optional environment variables:
| Variable | Purpose |
|---|---|
GIMP_CONSOLE |
Full path to gimp-console-3.exe if it is not auto-detected |
GIMP_MCP_BACKEND |
headless (default) or bridge |
GIMP_MCP_BRIDGE_PORT |
Bridge port, default 50472 |
Tools
Inspection
| Tool | Purpose |
|---|---|
gimp_status |
Check GIMP is reachable; reports both backends. Start here if something is wrong. |
inspect_image |
Dimensions, layers, orientation. Dimensions are as displayed. |
check_image_spec |
Validate against a dimension spec; pass/fail with measured dimensions and a plain-language reason. |
Single image
| Tool | Purpose |
|---|---|
crop_image |
Exact pixel rectangle. Rejects out-of-bounds rather than silently clamping. |
crop_square |
Largest square; anchor = center/top/bottom/left/right/corner. |
crop_to_aspect |
Target ratio (1.0 square, 1.3333 for 4:3, 1.7778 for 16:9), max area. |
resize_image |
By width, height, or max_edge. Aspect preserved by default. |
adjust_image |
Brightness/contrast, restricted to -0.5..0.5. |
fit_to_spec |
One shot: fix orientation by cropping, upscale to a minimum, downscale to a maximum, optional touch-up. |
process_image |
Custom operation pipeline in one pass (one JPEG re-encode). |
Batch
| Tool | Purpose |
|---|---|
batch_process |
Arbitrary pipeline over a folder. |
batch_fit_to_spec |
Conform a whole folder to one dimension spec. |
batch_check_image_spec |
Read-only audit; triage before editing. |
A whole batch runs inside one GIMP invocation. GIMP's console takes several
seconds to start, so spawning per file would be slow — measured at ~2.4x
cheaper per file for a small folder, and the saving grows with folder size. A
file that fails does not abort the run; it lands in errors and the rest
continue.
Live control (needs the bridge plug-in)
| Tool | Purpose |
|---|---|
live_list_images |
What is open in the running GIMP. |
live_screenshot |
Flattened snapshot of the canvas, so you can see and iterate. |
live_run_python |
Arbitrary Python in the live context; assign to result. |
live_stop_bridge |
Stop the bridge, leave GIMP open. |
Start the bridge in GIMP: Filters > Development > Start MCP Bridge.
Image specifications
check_image_spec, fit_to_spec and their batch equivalents share one spec
model. Every constraint is optional — 0 means no limit, and orientation
any means no orientation requirement.
| Field | Values |
|---|---|
min_width, min_height |
pixels, 0 for no minimum |
max_width, max_height |
pixels, 0 for no maximum |
orientation |
any, square, landscape, portrait, square_or_landscape, square_or_portrait |
fit_to_spec satisfies a spec in three ordered steps: crop to correct the
orientation, upscale to reach the minimum, downscale to respect the maximum.
Constraints already satisfied leave the framing untouched.
// A square image at least 1000x1000, capped at 2000x2000
{ "orientation": "square", "min_width": 1000, "min_height": 1000,
"max_width": 2000, "max_height": 2000 }
Colour adjustment is deliberately limited
adjust_image restricts brightness/contrast to -0.5..0.5 and rejects
anything outside it rather than clamping. Values beyond roughly ±0.15 visibly
change the character of a photo, which matters when an image needs to
represent a real subject faithfully. There is intentionally no saturation
boost or "auto enhance".
Verification
Run the suite:
python -m pytest tests/ -v
Tests that need real images are skipped unless you point them at some:
export GIMP_MCP_TEST_IMAGE=/path/to/photo.jpg # ideally EXIF-rotated
export GIMP_MCP_TEST_REFERENCE=/path/to/photo-square.jpg
GIMP_MCP_TEST_REFERENCE should be an independently produced centred square
crop of GIMP_MCP_TEST_IMAGE — cropped by hand in GIMP, for example. The
headline test asserts that crop_square reproduces that reference, rather
than merely running without error.
On the reference photo used during development (a 4000x3000 JPEG with EXIF orientation 6, displaying as 3000x4000):
crop_square vs hand-made reference : mean abs diff 0.236, max 18, outliers 0.0014%
same crop via the bridge backend : mean abs diff 0.236, max 18, outliers 0.0014%
That residual is JPEG re-encode noise — re-encoding alone gives ~0.5 mean — not a geometry difference, and both backends agree exactly.
The suite also covers displayed-orientation reporting, orientation and minimum-size specs, out-of-bounds crops being rejected, out-of-range adjustments being rejected, brightness moving pixels the right way, chained pipelines, aspect-ratio cropping, batch across a folder, the read-only audit, clear errors for missing files, and a full pass over the real MCP stdio protocol.
Troubleshooting
gimp-console not found — set GIMP_CONSOLE to the full path of
gimp-console-3.exe.
Bridge tools fail with "Could not reach the GIMP bridge" — GIMP is not
open, or the bridge was not started. Run Filters > Development > Start MCP
Bridge. gimp_status shows both backends at once.
The menu item is missing after installing — restart GIMP; it only scans
plug-ins at startup. Confirm the layout is
plug-ins/gimp-mcp-bridge/gimp-mcp-bridge.py (the folder name must match the
file name).
Diagnosing the plug-in — a GIMP plug-in is a separate process whose stderr
is invisible when GIMP runs as a GUI app on Windows. The bridge writes to
bridge.log next to the installed plug-in.
A colour-profile dialog blocks GIMP on startup when opening an image with an embedded profile in GUI mode. It does not appear in headless mode, which is another reason batch work uses the headless backend.
Batch timed out — the default is 600s for the whole run; very large folders may need more.
Known limitations
- Live control is only lightly exercised. It is verified working (open an image, list, screenshot, edit live, and crop through the bridge with output identical to headless), but it has had far less use than the headless path. Treat headless as the trustworthy one.
- The bridge executes arbitrary Python by design. It is loopback-only and started manually rather than automatically, but anything that can reach localhost on the machine can drive GIMP while it is running. Stop it when not in use.
- Bridge start blocks its own plug-in process — that is what keeps it alive. It does not freeze GIMP's UI, but GIMP shows the plug-in as running.
- The GUI menu item itself is not automated-test covered. The procedure it invokes is verified; the click path is not.
- Only Windows is verified. The code paths are cross-platform and the installer handles Linux/macOS config directories, but neither has been tested.
- The
mcp2.x SDK is not supported yet -- see the note under Install. - No AI background removal or style transfer. Some comparable projects advertise these without a working implementation behind them; they are deliberately not claimed here.
Notes on prior art
The split between a GIMP-side plug-in exposing a bridge and a standalone MCP server process that connects to it as a client is a natural shape for this problem and is used by other GIMP MCP projects. Batch processing and preset-style pipelines are common to several. Live-canvas control exists elsewhere via D-Bus, replaced here with loopback TCP for Windows support. No code was copied from any of them; the Windows specifics — the real plug-in path, the plug-in process lifetime, the run-callback signature, and the EXIF behaviour — were established directly against GIMP 3.2.4.
License
MIT — see LICENSE.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
E2B
Using MCP to run code via e2b.