Leash is a private, offline personal AI that lives across devices you own. Every inference, embedding, retrieval, vision, speech, delegation, and fine-tuning path runs through QVAC on local hardware. There is no setting that swaps in a cloud model. Leash won first place overall and the Build in Public prize at QVAC Hackathon I, “Unleash Edge AI” (July 2026).

Source code

Apache-2.0, every layer real.

Leash docs

Install, platforms, and the proofs.

Website

The product surface.

​
Five layers, one organism

Leash is the product you talk to. Mycelium is the runtime underneath it.

​
What it measured

These numbers come from the committed evidence bundle in the repository, on the hackathon hardware.

​
Lessons you can take without building Leash

  1. Give the model a bounded job. A local model does not need to beat frontier models. It needs a clear responsibility, private context, deterministic guardrails, and review.
  2. Make privacy a routing rule, not a promise. Leash’s router never sends a private prompt to a public peer, and nothing can override that.
  3. Ship receipts. Every claim in the README links to a log, a smoke script, or a JSON summary. If you say it runs offline, prove it.
  4. Contribute upstream. Building Leash surfaced a gap in the SDK, which became a contribution to QVAC itself: tetherto/qvac#2459, image_url content in chat completions.

Build these ideas yourself

Every Bring AI Home project applies the same principles to one everyday app.