OpenClaw Draw Things

Local AI image generation for OpenClaw using Draw Things CLI on Apple Silicon. Generate images with Stable Diffusion, FLUX, and other models without API costs.

Features

• 🖼️ Local image generation - No API keys needed
• 🎨 Image editing (img2img) - Transform existing images with AI
• 🍎 Apple Silicon optimized - Uses Core ML for fast inference
• 🤖 Multiple models - SD, FLUX, and more
• ⚡ OpenClaw integration - Native tool support

Prerequisites

• macOS with Apple Silicon (M1/M2/M3/M4)
• Draw Things CLI installed via Homebrew:

  brew tap drawthingsai/draw-things
  brew install draw-things-cli
  

• (Optional) Draw Things App from Mac App Store — easier GUI for browsing and downloading models
• OpenClaw 2026.4.0 or later
• AI Models downloaded (see Models section below)

💡 Note: Only the CLI is required. The Draw Things app is optional but provides a convenient GUI for browsing and downloading models.

Installation

From ClawHub (Recommended)

openclaw plugins install openclaw-draw-things --dangerously-force-unsafe-install

⚠️ Security Notice: This plugin executes shell commands to run the Draw Things CLI. OpenClaw's security scanner flags this as "dangerous code patterns". This is expected and necessary for the plugin to function — it uses child_process to spawn the draw-things-cli binary for local image generation. Use the --dangerously-force-unsafe-install flag to proceed with installation.

Manual Installation

git clone https://github.com/acwilan/openclaw-draw-things.git
cd openclaw-draw-things
npm install
npm run build
openclaw plugins install "$(pwd)"

Configuration

Add to your ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "draw-things": {
        "enabled": true,
        "config": {
          "cliPath": "draw-things-cli",
          "outputDir": "~/Downloads/draw-things-output",
          "defaultModel": "realistic_vision_v5.1_f16.ckpt",
          "defaultSize": "1024x1024",
          "defaultEditStrength": 0.5,
          "defaultPromptMode": "auto",
          "defaultPromptAppend": "watercolor, soft brush strokes, cohesive painterly style",
          "defaultConfigJson": {
            "sampler": 12,
            "loras": [
              {
                "mode": "all",
                "file": "breath_of_the_wild_style_lora_f16.ckpt",
                "weight": 1
              }
            ]
          }
        }
      }
    }
  },
  "agents": {
    "defaults": {
      "imageGenerationModel": {
        "primary": "draw-things/realistic_vision_v5.1_f16.ckpt"
      }
    }
  }
}

Config Options

Models

The provider now uses a generated Draw Things CLI model catalog plus curated overrides for defaults like size, steps, CFG, prompt mode, and high-resolution behavior. Unknown model names still work via best-effort type inference, but adding curated overrides in src/model-metadata.ts gives more predictable behavior.

At runtime, the plugin checks draw-things-cli models list --downloaded-only and falls back to a downloaded model if the configured/requested model is not present locally.

Models are stored in:
~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models/

Or specify a custom modelsDir in the plugin config.

Downloading Models via CLI

If you prefer not to use the Draw Things app, you can download models entirely via CLI:

# List currently downloaded models
draw-things-cli models list --downloaded-only

# List all available models for download
draw-things-cli models list

# Download a specific model
draw-things-cli models ensure --model <model-name>

💡 Tip: If using a custom models directory, add --models-dir <path> to match your plugin config.

Recommended Models

Via Draw Things App (optional GUI method):
Open Draw Things app → Models → Download Models

Usage

Once installed and configured, OpenClaw can generate images:

Generate an image of a sunset over mountains

Or use explicit tool calls:

Use image_generate to create a cartoon cat

Default Prompt Append

Use defaultPromptAppend to lock a Draw Things-backed agent into a consistent visual profile. The plugin appends this text to every generation/edit prompt before model-aware prompt optimization, so a user prompt like a portrait of a woman can become a portrait of a woman, watercolor, soft brush strokes, cohesive painterly style.

Example profiles:

{
  "defaultPromptAppend": "watercolor, soft brush strokes, cohesive painterly style"
}

{
  "defaultPromptAppend": "cartoon illustration, bold outlines, vibrant colors"
}

Draw Things Configuration Overrides and LoRAs

Use defaultConfigJson for Draw Things-native options that are not part of OpenClaw's shared image-generation contract. The object is serialized and passed to draw-things-cli generate --config-json for both initial generation and high-resolution generation passes.

{
  "defaultConfigJson": {
    "sampler": 12,
    "hiresFix": false,
    "loras": [
      {
        "mode": "all",
        "file": "breath_of_the_wild_style_lora_f16.ckpt",
        "weight": 1
      }
    ]
  }
}

The CLI applies these values before the plugin's explicit model, dimensions, steps, CFG, and edit-strength arguments, so the shared OpenClaw request controls continue to win for overlapping settings.

OpenClaw's current image_generate request only defines provider-specific options for OpenAI. This means Draw Things overrides are persistent plugin configuration, not per-call tool parameters; supporting dynamic per-request LoRA selection requires an upstream extension of the image-generation contract.

Image Editing (img2img)

Transform existing images using AI:

Turn this photo into a watercolor painting [attach image]

Or explicitly:

Edit this image to look like a comic book style

The plugin supports:

• One input image per edit request
• Configurable strength via defaultEditStrength (how much the image changes)
• Style transformations - watercolor, oil painting, cartoon, etc.
• Detail enhancement - improve or modify specific aspects

Tips for best results:

• Use clear style descriptions (e.g., "oil painting", "anime style", "sketch")
• For subtle changes, the strength is moderate (0.5)
• Works with any supported model

Troubleshooting

Provider does not show as configured

The plugin reports configured only when it can find draw-things-cli, the output path is plausible, and draw-things-cli models list --downloaded-only returns at least one local model.

"No image-generation provider registered"

Make sure the imageGenerationModel.primary is set correctly:

"imageGenerationModel": {
  "primary": "draw-things/your-model.ckpt"
}

"Model not found"

• Verify the model file exists in Draw Things
• Check the exact filename in the Models directory
• Ensure modelsDir config matches your setup if using custom location

"draw-things-cli command not found"

Install the CLI via Homebrew:

brew tap drawthingsai/draw-things
brew install draw-things-cli

Plugin not loading

Check the plugin is enabled:

openclaw plugins list

If needed, restart the gateway:

openclaw gateway restart

Development

# Install dependencies
npm install

# Build
npm run build

# Watch mode
npm run dev

# Run tests
npm test

Releasing

# Patch version (1.0.0 → 1.0.1)
npm run release:patch

# Minor version (1.0.0 → 1.1.0)
npm run release:minor

# Major version (1.0.0 → 2.0.0)
npm run release:major

This handles version bump, manifest sync, changelog, commit, tag, and push automatically.

License

MIT © Andres Rovira

Links

• GitHub
• Draw Things
• Draw Things CLI
• OpenClaw
• ClawHub Package