Zygos
Local-First AI Runtime
Run AI on your terms: private or public, and always inspectable.
Zygos is an open, local-first AI runtime you host yourself. Point it at a private local model or a public provider, and see which one handled every request. Memory and history stay in a local database on your machine, and nothing phones home.
It is built to be inspected: the model, memory, and tools behind each answer are visible, so you can verify what was used and what left your machine.
The problem
AI increasingly handles private conversations, documents, long-term memory, and access to real tools. Too often that means sending sensitive context to services you can’t see into, accepting agent behavior you can’t audit, and getting locked into a single provider.
Why Zygos
Private or public models
Run locally with Ollama or vLLM, or connect OpenAI or Anthropic when you choose, with every request showing which model handled it.
Your data stays put
Memory and history live in a local database on your own machine or VM. There is no managed platform in the middle.
Auditable, not opaque
Inspect the model route, memory, and tools behind an answer through the runtime manifest and inspection panels.
Real tools, gated by you
Zygos can read and write files, fetch the web, and run commands, but side-effecting tools ask before they act.
Improvement you approve
Zygos can propose new skills and refinements, but never changes its own behavior without human review and testing.
Local voice
Speech in and out on your own machine, with hands-free conversation and the ability to interrupt a reply mid-sentence. Opt-in.
No telemetry and no lock-in: Zygos ships no analytics, and models, memory backends, tools, and providers can be swapped without rebuilding.
Status
- v1
- Stable TypeScript command-line runtime with local and public model support, local storage, and a permissioned tool framework. Frozen; bug fixes only.
- v2
- In active development: a self-hosted Python server with a live chat-and-tools loop and a React web app with chat, inspection panels, and local voice.
- Next
- Human-approved skill learning, workflows, a scheduler with human-in-the-loop guards, and a single-command installer.
- License
- Apache 2.0