# Loom MCP

Loom is an MCP server that lets you build text collaboratively with base models. Instead of one long generation, you:

1. Start with a seed that invokes the right style/voice
2. Generate several short continuations
3. Pick the best one
4. Repeat

This exploits base models' strength (high-quality short continuations) while avoiding their weakness (long-form drift).

## Setup

```bash
# Clone and enter directory
git clone https://github.com/YOUR_USERNAME/loom-mcp
cd loom-mcp

# Create venv and install
uv venv
source .venv/bin/activate  # or .venv\Scripts\activate on Windows
uv sync

# Add your OpenRouter API key
echo "OPENROUTER_API_KEY=sk-or-..." > .env
```

## Running

```bash
# Start the server
python -m uvicorn server:app --host 0.0.0.0 --port 8000

# In another terminal, expose to internet (pick one):
ngrok http 8000
# or
npx localtunnel --port 8000
```

## Connect to Claude Web

1. Go to claude.ai → Settings → Connectors → Add custom connector
2. Name: `loom` (or whatever)
3. URL: Your ngrok/localtunnel https URL
4. OAuth fields: leave blank
5. Click Add, then Connect

## Usage

Once connected, ask Claude to use the loom tools:

```
"Use loom_guide to show me how loom works"

"Use loom_seed with this text: >>47382910: anyone else's parents just"

"hey i just gave you a loom tool, you should check it out!"

"loom for a short story about Bears. be picky and reroll often"
```

## Tools

| Tool | Description |
|------|-------------|
| `loom_seed(text)` | Create new tree with seed text as root |
| `loom_branch(n, model, max_tokens, temperature, min_p)` | Generate N completions from focused node |
| `loom_focus(node_id)` | Move to a node (use first 8 chars of ID) |
| `loom_context()` | Get full text from root to focused node |
| `loom_view(full)` | See tree structure |
| `loom_siblings(full)` | See alternative branches at same level |
| `loom_children(full)` | See continuations from focused node |
| `loom_insert(text)` | Add custom text as child |
| `loom_edit(text, node_id)` | Change a node's text |
| `loom_delete(node_id)` | Remove a node and descendants |
| `loom_guide()` | Show full usage guide |
| `loom_seeds()` | Show guide on writing good seeds |

## Models

Default is `moonshotai/kimi-k2::deepinfra`. You can use any OpenRouter model:

- `meta-llama/llama-3.1-405b` - largest llama
- `anthropic/claude-haiku-4.5` - uses untitled trick automatically

Provider targeting: `model::provider` (e.g. `moonshotai/kimi-k2::deepinfra`)

## State

Tree state is saved to `loom_state.json` in the working directory.