Use case

A private AI assistant you actually own.

ThreadWeaver runs on the Raspberry Pi 5 and gives you a full-featured chat interface — conversation branching, model switching, provider routing — all pointed at models running on your own Jetson Orin Nano. Nothing leaves your network.

Most self-hosted chat setups are fragile: a Docker container, a model you pulled last quarter, no way to switch providers or compare responses. Claw ships a complete, pre-integrated stack that works day one and stays working when you pull a new model or update ThreadWeaver.

What's included

  • ThreadWeaver — browser-based chat UI with conversation branching, model comparison, and operator console. Runs on the RPi5, proxied through Caddy with TLS.
  • Ollama + CUDA inference — models run on the Jetson's GPU, not your laptop. 24/7 availability without keeping a workstation on.
  • LiteLLM proxy — one endpoint for all providers. Switch between local Ollama models and cloud APIs without changing client config.
  • OpenClaw agent runtime — give your assistant real tools: file access, web search, shell exec, scheduling, LED feedback. It can do things, not just talk about them.

Open your chat from any device

Once installed, ThreadWeaver is available on your local network at https://claw.local from any browser — phone, tablet, laptop. The Caddy reverse proxy handles TLS with a self-signed CA you trust once. 

# First time only: trust the cluster CA
curl http://claw.local/ca.crt -o claw-ca.crt
# Install in your browser or OS trust store

# Then just open: https://claw.local

Which model to start with

For general chat, we recommend granite4.1:8b as the default — it scored 10/10 on our tool-call benchmark and gives reliable, structured responses. For more natural conversation, qwen2.5:7b is the better pick. 

# Pull a model on the Jetson
ssh picocluster@clustercrush
ollama pull granite4.1:8b

# Or pull from the ThreadWeaver model panel — no SSH needed

Adding tools to your assistant

OpenClaw connects to MCP tool servers for real-world capabilities. The Claw ships with tools for system monitoring, file operations, web search, and LED control. Adding more is a config edit: 

// In openclaw.json — add any MCP-compatible tool server
{
  "mcpServers": {
    "weather": {
      "command": "uvx",
      "args": ["mcp-server-fetch"]
    }
  }
}

See the MCP tools reference for the built-in tools and how to add your own.

Experimenting with models

One of the advantages of owning the hardware: you can run your own benchmarks. We publish scripts and methodology for testing model capability across 30 real-world tasks — web research, scheduling, multi-tool chains, natural-language command interpretation.

See SMEEP benchmark results for what we've measured, and the scripts to reproduce it on any Ollama-capable hardware.

Next: 5-minute quickstart  ·  Benchmark results  ·  MCP tool reference