wool

A tree-based completion explorer for base LLMs. Generate multiple completions simultaneously, branch from any node, and navigate through possibility space on an infinite canvas. Flask backend with vanilla JavaScript frontend - no setup friction, just branches and text.

Quick Start

1. Install Python (3.7 or higher) if you haven't already
2. Install the required packages:

   pip install flask requests
   

3. Run the application:

   python app.py
   

4. Open your browser and go to http://127.0.0.1:5000

First Time Setup

1. Launch the server and browser as above (see Quick Start)
2. The Settings panel should be open on the left side
3. In the Model/Endpoint field, enter:
   • An OpenRouter model: anthropic/claude-sonnet-4.5 or deepseek/deepseek-r1-0528
   • A local/remote server: http://localhost:1234/v1 or https://your-tunnel.trycloudflare.com/v1
4. Enter your API key (if using OpenRouter)
5. Adjust temperature, min_p, and max_tokens as desired

Creating and Exploring Trees

Starting a Tree:

• Enter seed text in the bottom textbox
• Click "Create Root" to generate your first node

Generating Branches:

• Click any node to focus it (dark border indicates focus)
• Click the + button to generate child nodes
• Adjust "N Siblings" in settings to generate multiple alternatives at once
• All siblings generate in parallel

Navigation:

• Click and drag the canvas to pan
• Scroll to zoom in/out
• Click any node to refocus and see its context

Visibility:

• The focused node is always visible
• Ancestors up to root are visible (your current path)
• Siblings at each level are visible (alternative branches)
• Immediate children of focused node are visible (next steps)
• Everything else is hidden to reduce clutter

Saving/loading:

• Creating a new seed will wipe the entire tree, so be careful!
• Individual leaves have a "download as .txt" when viewed
• Download a whole tree as json
• Upload trees from json files saved with wool

Model Settings

Temperature: Controls randomness. Higher = more creative/varied, lower = more focused/repetitive.

Min P: Minimum probability threshold. Higher = more conservative token selection, lower = wider vocabulary.

Max Tokens: Length of each completion. Adjust based on how much text you want per node.

N Siblings: Number of parallel completions to generate from the focused node. 1-8 recommended.

Endpoints

OpenRouter:

• Enter any model in provider/model-name format: anthropic/claude-sonnet-4.5, deepseek/deepseek-r1-0528, meta-llama/llama-3.1-405b
• Get your API key from https://openrouter.ai/settings/keys
• Add it in the settings panel under "API Key"

Provider Targeting (:: syntax):

• Force a specific backend on OpenRouter: moonshotai/kimi-k2::deepinfra/fp4
• Format: model::provider - useful when you want a specific host for speed/quality/cost
• The provider after :: must be available for that model on OpenRouter

OpenAI-Compatible Servers:

• Enter any URL ending in /v1: http://localhost:1234/v1, http://192.168.1.100:8080/v1, https://your-tunnel.trycloudflare.com/v1
• Works with LM Studio, vLLM, llama.cpp server, text-generation-webui, tabbyAPI, etc.
• Can be local (localhost) or remote (LAN IP, cloudflare tunnel, ngrok, etc.)
• No API key needed for most local servers

Auto-detection:

• Starts with http:// or https:// → OpenAI-compatible server
• Everything else → OpenRouter model (invalid models will error when you generate, not when you enter them)
• Important: OpenAI-compatible URLs must end in /v1 or the request will fail

Tree Persistence

Tree state is automatically saved to tree_state.json on the server. Your exploration survives page refreshes, server restarts, and browser crashes. One tree per server instance.

Contributing

Contributions are welcome! Feel free to submit issues and pull requests.