GIMP Agent MCP

Let your AI agent edit in GIMP. Keep the layers. See what changed.

Connect Claude, Codex, Cursor or another MCP client to GIMP 3. The agent can work in your open window, inspect the result, refine an edit and save an XCF you can keep working on. Headless mode runs the same workflows over folders of images.

See it working

Watch the 20-second MP4 · Open the editable XCF · Reproduce the MCP session

This is a captured-step walkthrough of the real GIMP window, with pauses shortened. Three text layers are changed through MCP in one batch. One Ctrl+Z restores all three; Ctrl+Y brings them back. The text, artwork and background stay on separate named layers. This demonstrates actual tool execution, not a claim that every model will follow the same plan from a prompt.

Try asking

“Change the headline and supporting text in this card. Keep them as text layers, group the edits into one undo step, and show me the result before exporting.”

“Make this transparent artwork into a Telegram sticker with a white outline and shadow. Save a 512 × 512 PNG.”

“Pack these equal-sized PNG frames into a sprite sheet. Keep an XCF master, write the atlas JSON and verify the exported pixels against the originals.”

The included skills guide layered artwork, sprite sheets and batch jobs. They tell the agent to inspect and measure the actual output before calling a job complete.

Install

You need GIMP 3.2.x, Python 3.11+, uv and an MCP client. GIMP must be installed on the same machine as the server. Start GIMP once so its user profile exists.

uvx gimp-agent-mcp install-plugin
uvx gimp-agent-mcp install-skills
uvx gimp-agent-mcp doctor

Restart GIMP after installing or upgrading the bridge. In your open GIMP window, choose Filters → Development → Start Agent Bridge.

Add this configuration to your client's MCP settings:

{
  "mcpServers": {
    "gimp": {
      "command": "uvx",
      "args": ["gimp-agent-mcp", "serve"]
    }
  }
}

For Claude Code:

claude mcp add gimp -- uvx gimp-agent-mcp serve

Claude Code can also install the server and skills as a plugin; see skills installation. For AI background removal, install the optional model runtime with uvx --from "gimp-agent-mcp[segmentation]" gimp-agent-mcp serve (use the same arguments in your MCP config). The first use downloads a model; image processing stays local.

Working from source? Run uv sync --extra dev, then uv run gimp-agent-mcp install-plugin. Configure the server as uv run --no-sync --directory /absolute/path/to/repo gimp-agent-mcp serve. --no-sync avoids Windows executable-lock errors when another server is running.

Work in your window, or run a batch

• Your existing window: start the bridge from the menu. Continue editing the same document by hand.
• A shortcut: uvx gimp-agent-mcp shortcut creates a launcher that starts GIMP with the bridge enabled.
• Agent-launched: gimp_launch(mode="gui") opens GIMP; mode="headless" runs without a window.

gimp_status and gimp_list_images report the image attached to GIMP's active user display, its selected layer ids, selection bounds and dirty state. When the user says "this layer" or "the selection", read that context before editing. Use gimp_context(image_id) for a specific document or to bring it forward with present=true. If multiple bridges run, gimp_status follows the newest bridge file; isolate test profiles when a user session is active.

What makes this useful

Other GIMP MCP projects also provide TCP bridges and visual feedback. This project's emphasis is editable output, measurement, grouped edits and reproducible validation. Generic API access is broad, but does not guarantee that every GIMP procedure or GEGL operation works with every argument combination.

Tools (40)

Images and items use integer IDs. Colours accept hex, common names, CSS RGB strings or component arrays; enums use nicks returned by the describe tools. Call gimp_help("batch") for grouped-edit examples. gimp_run_python may be disabled, reducing the tool count by one.

Common work no longer needs procedure discovery. For example, use gimp_adjust(layer_id, "brightness_contrast", {"brightness": 0.15}), gimp_canvas(image_id, "crop", width=1200, height=630, x=40, y=20), or gimp_draw(image_id, layer_id, "line", color="#ffffff", x=20, y=40, x2=300, y2=40, line_width=6). Adjustments default to editable layer effects where GIMP supports them; canvas and drawing actions bake pixels or structure. Every result identifies the runtime operation selected before execution.

For visual placement, ask for a diagnostic preview without marking the working image:

gimp_render(image_id=3, overlay=["grid", "layers", "selection"], grid_size=100,
            points=[{"x": 600, "y": 340, "label": "headline centre"}])

Recipes

Brand kits for compose

Store a brand's colours, fonts, assets and canvas sizes once, then refer to them by token. Profiles are JSON files in the per-user config directory (%APPDATA%\gimp-agent-mcp\brands on Windows, ~/Library/Application Support/gimp-agent-mcp/brands on macOS, $XDG_CONFIG_HOME/gimp-agent-mcp/brands elsewhere; GIMP_AGENT_BRANDS_DIR overrides). Font names are checked against the running GIMP when a profile is saved.

gimp_brand(action="set", profile={"name": "example",
  "colors": {"bg": "#0f1117", "accent": "#1f3a8a", "text": "#ffffff"},
  "fonts": {"headline": "Bahnschrift Bold"}, "assets": {"logo": "/brand/logo.png"},
  "canvases": {"x-card": [1200, 675], "og": [1200, 630]}})

gimp_run_recipe("compose", {"output_path": "out/card.png", "manifest": {
  "brand": "example", "canvas": "x-card", "background": "@bg",
  "items": [{"type": "image", "path": "@logo", "x": 40, "y": 40, "width": 160},
            {"type": "text", "text": "Launch day", "font": "@headline", "size": 72, "color": "@text", "x": 40, "y": 240}]}})

Tokens are resolved in the server before anything reaches GIMP, so a token manifest becomes exactly the manifest you would have written by hand. Literal values still work and win over the profile; an unknown token fails with the profile name and its available keys. gimp_help("compose") has the full rules.

Recipe arguments and examples · Asset workflows · Skills

Tested scope and limits

Version 0.5.0, beta. Tested with Windows GIMP 3.2.4, Linux GIMP 3.2.2 (Ubuntu 26.04), and macOS 15 GIMP 3.2.4 on Apple Silicon and Intel. Each platform runs the full smoke suite, targeted capability proof and all 23 everyday editing actions against real GIMP. GIMP 2.10 is unsupported; earlier 3.x releases are not part of the current test matrix.

• gimp_edit_batch accepts bounded layer, text, selection, path, mask and filter edits. gimp_undo_group(image_id, action="start", name="agent: sticker pass") holds a group across calls; end it with action="end". gimp_status lists open groups. A client disconnect closes its groups. The supplied name is tracked by the bridge, but GIMP 3's public undo-group API cannot set the Undo History label. The human uses GIMP's Undo/Redo; there is no programmatic undo endpoint.
• uv run python scripts/undo_group_smoke.py checks three edits against a pre-turn snapshot after you press Ctrl+Z once in a GUI GIMP session. The headless gimp-agent-mcp smoke check covers group lifecycle and disconnect cleanup.
• Some GEGL source operations, including linear gradients, are not drawable filters. Use the PDB gradient-fill procedure instead. Vector warp/liquify parity is not claimed; see the warp investigation note.
• Sprite packing takes equal-sized frames (sprite_sheet_pack, with margin and spacing) or mixed-size PNGs and open layers (sprite_atlas_pack, with padding and trimming). Neither aligns animation or rotates sprites.
• Long work reports progress and can be cancelled with gimp_cancel at checkpoints: between recipe items, batch files and edit-batch steps, between passes of gimp_run_python code that calls checkpoint(), and before a filter starts. A single GEGL merge already running cannot be interrupted, and GIMP's own per-procedure progress is not forwarded yet.

Verify your setup

uvx gimp-agent-mcp smoke
# From the source checkout:
uv run pytest
uv run python scripts/capability_proof.py /path/to/proof-output
uv run python scripts/everyday_proof.py /path/to/proof-output

CI checks supported Python versions, and separate jobs exercise real GIMP on Windows, Linux and both macOS architectures. Release publication depends on all three real-GIMP workflows passing. The general capability proof measures curves, gradients and brush output and checks batch recovery, document boundaries, diagnostic overlays, nested previews and snapshot cleanup. A second proof resolves and exercises every action in gimp_adjust, gimp_canvas and gimp_draw, recording dimensions, layer counts and measured pixels. See validation details.

Each smoke and proof script runs on a GIMP it launched itself and, when it shuts that GIMP down, waits until the bridge is really gone (port closed, bridge file removed, plug-in process exited) before returning. A bridge that has been asked to shut down refuses new connections and new requests at once. The smoke asserts this in its shutdown (bridge gone before return) check, and CI sets GIMP_AGENT_REQUIRE_OWN_GIMP=1 so a proof that finds an earlier step's bridge still answering fails with that reason instead of borrowing it. This fixed the macOS arm64 live-smoke failures of 15 October 2026, which looked like GIMP crashes but were a proof attaching to the previous step's GIMP while it quit. The live jobs do not retry checks: if the macOS job fails, its log says whether macOS wrote a GIMP crash report, and the macos-gimp-proof-<arch> artifact holds the proof reports, the GIMP launch log and any crash reports.

Architecture and trust

MCP client → stdio server → authenticated loopback TCP → plug-in inside GIMP

The plug-in runs operations on its GLib main loop, without worker threads. It uses a per-install token and only binds to 127.0.0.1. PDB calls, recipes and Python can execute code with your user permissions: connect only trusted clients. Disabling gimp_run_python is not a sandbox. Security · Architecture

Clean-room implementation under Apache-2.0; no code copied from other GIMP MCP projects. Roadmap · Changelog · Report a bug