Use GroomLab from AI agents

GroomLab has no chat window of its own. Instead, groomlab-cli has an MCP server built in, so the AI app you already use can drive GroomLab directly. MCP (the Model Context Protocol) is an open standard for connecting AI apps to tools: the app starts a small program on your machine, asks it which tools it has, and lets its model call them. With GroomLab connected, you can ask for "a curly bob on Nika, a bit redder", and the agent builds the groom, looks at a preview, adjusts it and saves a .groom file that Maya, the app, Blender and the render procedurals open.

The server runs on your machine and makes no network connections. It reads and writes files only inside one folder that you choose.

Where groomlab-cli isCopy Anchor Link

groomlab-cli comes in the Maya download, in the GroomLab\bin folder of the module. If you unzipped it to Documents\GroomLab, that is:

C:\Users\<you>\Documents\GroomLab\GroomLab\bin\groomlab-cli.exe

On Linux it is GroomLab/bin/groomlab-cli in the groomlab-<version>-linux-x86_64.tar.gz archive. The standalone app's zip does not include it.

The server is started as groomlab-cli mcp --root <folder>. The root is the folder the agent works in: meshes it grooms on, and every .groom, .abc, .usda and .png it writes. Without --root, the server uses the folder it was started in. Make a folder for it, for example Documents\grooms.

Connect your AI appCopy Anchor Link

Each app has its own config file. In the snippets below, replace the paths with your own. In JSON, backslashes are written twice.

Claude DesktopCopy Anchor Link

Open Settings > Developer > Edit Config, add GroomLab to claude_desktop_config.json, then restart Claude Desktop:

{
  "mcpServers": {
    "groomlab": {
      "command": "C:\\Users\\you\\Documents\\GroomLab\\GroomLab\\bin\\groomlab-cli.exe",
      "args": ["mcp", "--root", "C:\\Users\\you\\Documents\\grooms"]
    }
  }
}

Claude CodeCopy Anchor Link

Run this once, in a terminal:

claude mcp add --scope user groomlab -- "C:\Users\you\Documents\GroomLab\GroomLab\bin\groomlab-cli.exe" mcp --root "C:\Users\you\Documents\grooms"

/mcp in Claude Code shows whether it is connected. Leave out --root to work in the folder Claude Code was started in.

CursorCopy Anchor Link

Add the server to ~/.cursor/mcp.json (every project) or .cursor/mcp.json (one project):

{
  "mcpServers": {
    "groomlab": {
      "command": "C:\\Users\\you\\Documents\\GroomLab\\GroomLab\\bin\\groomlab-cli.exe",
      "args": ["mcp", "--root", "C:\\Users\\you\\Documents\\grooms"]
    }
  }
}

VS CodeCopy Anchor Link

Add it to .vscode/mcp.json in your workspace (or run MCP: Add Server from the command palette), then use it from Copilot's agent mode:

{
  "servers": {
    "groomlab": {
      "type": "stdio",
      "command": "C:\\Users\\you\\Documents\\GroomLab\\GroomLab\\bin\\groomlab-cli.exe",
      "args": ["mcp", "--root", "${workspaceFolder}"]
    }
  }
}

Any other app that supports MCP servers over stdio works the same way: the command is groomlab-cli, and the arguments are mcp --root <folder>.

The toolsCopy Anchor Link

The agent sees these tools, each with a description and a JSON Schema for its arguments. Grooms are held in memory by name, so one session can work on several grooms. Each tool that changes a groom answers with a short summary: strand and point counts, guides, bounds, and any node's errors.

An example sessionCopy Anchor Link

You: Make tube-lock anime hair on Nika and show me.

A typical agent does something like this:

1. list_presets, then new_groom with {"name": "anime", "mesh": "nika", "preset": "tube_locks"}. The summary has an error: Scatter's guide radius needs guides.
2. sample_scalp with {"min_up": 0.3} for points on the top of the head.
3. add_guides with roots from those points, for example {"root": [0, 13, 0], "length": 18, "droop": 0.8}, and a ribbon tube for all of them: {"shape": "ribbon", "width": [1.6, 1.2, 0.2], "aspect": 0.35}.
4. render_preview with {"views": ["front", "side", "back", "three_quarter"]}. The agent looks at the image and changes what it sees: set_tube for flatter locks, set_params on clump.
5. save_groom with {"path": "anime.groom"}, which you then open in Maya or the app to keep grooming with brushes.

If a change goes wrong, the agent can call undo. If an argument is wrong, the tool's error says why (a param's range, the node names it has, a path outside the folder), and the agent tries again.

Groom recipesCopy Anchor Link

Typical graphs, as the shipped presets build them. Lengths are in scene units (centimeters on Nika, a head about 35 units tall). A preset applied to a scalp of another size is scaled to fit it.

• Natural hair (hair_basics): Scalp > Scatter (density 80, length 15) > Interpolate (with Guides) > Clump (clumpBy size, size 1.2, strength 0.8) > Noise (amplitude 0.05) > HairColor (melanin 0.7: 0 blond, 1 black; redness) > Width (width 0.007, tapered profile)

  Resample (16 segments) > Output. Guides do most of the styling: without them, strands grow straight out along the scalp's normals.

• Curly (curly): as above, with Curl after Clump (radius 0.45, frequency 0.4, coherence clump, so each clump curls as one ringlet), and Resample at 40 segments for smooth curls.
• Tube locks, anime and cartoon (tube_locks): Scatter with guideRadius 1 (hair only on the guides' tubes, fed the guides), Interpolate with fillTubes on (strands fill each guide's tube to its tip; tubeBlend softens the edge), Clump with followTubes. The tubes' shape, width, aspect and twist (set_tube) set each lock's look.
• Fur (creature_fur): Scalp > Scatter (density 25, length 3.5) > Length (uneven ends) > Flow (fur lies along painted direction arrows; without arrows, along its own direction) > Clump (small tufts) > Noise > HairColor (tipColor and tipStrength lighten the tips) > Width > Resample (8 segments). Made at wolf size (100 units).
• More looks: braids, wet, frizz, flyaways, gravity, two_tone_tips, parted_style and the feather and scale presets. list_presets describes each one.

Key params by node: Scatter density (strands per square unit) and length; Clump size or count, strength and profile (root to tip); Curl radius and frequency; Noise amplitude and frequency; Length fraction and randomness; HairColor melanin, redness, dye and the tip color; Width width and profile. describe_node gives every param with its range.

SafetyCopy Anchor Link

• One folder. The server reads and writes only inside its root. Paths are resolved first, so .. and links cannot leave it. This also applies to the files a groom names (caches, texture maps in params) and to save_to. GroomLab's own presets, nodes, sample meshes and docs are read where they are installed.
• No network. The server talks only to the app that started it, over stdin and stdout.
• Limits. Mesh files up to 512 MB, 5,000 guides per call, scripts up to 64 KB, previews up to 1024 pixels wide, exports up to 1,000 frames.
• Nothing is overwritten by surprise. Changes stay in memory until save_groom, save_node or export writes a file, and those writes go only to the paths the agent names, inside the root.
• Same calls, same results. Grooms evaluate deterministically, so the same calls give the same strands and the same preview pixels. Only timings differ between runs.

SchemasCopy Anchor Link

JSON Schemas for GroomLab's file formats, for agents and tools that write them directly: groom.schema.json (.groom, its top level), graph.schema.json (a graph, as grooms and presets hold it) and glnode.schema.json (compound and script nodes). See Groom file format for every field.

Not yetCopy Anchor Link

• A live bridge to Maya or the app. The server works on grooms in its own memory. To continue with brushes, open the saved .groom in Maya or the app.
• Painting maps. A preset that reads painted maps (density, length) says which maps are missing. These maps can only be painted in GroomLab itself.
• Flow arrows, parts and sculpts. These come from the brushes in Maya, the app and Blender.