Metis

A personal AI operating system that runs agents as a governed, multi-session workforce across two machines. Specialized lanes, persistent memory, task governance, and a control-center UI keep the system compounding instead of starting cold every session.

Specialized local-model lanes (generation, research, review, governance), steered by one orchestrating model

Machines coordinated; an always-on host plus an on-demand peer sharing one model host

700+

Governed tasks completed through the forward-only state machine

Local-first

Private MLX inference on-device; memory and state never leave the machines

FSL-1.1

Source-available framework published for others to vendor and run

Context

What agents need beyond a chat window

Useful agents need more than a chat window. They need memory that compounds, task coordination, live system visibility, safety gates, and a way to operate across machines without stale sessions corrupting shared work.

Approach

A governed, local-first operating layer

I built a local-first operating layer with specialized agent lanes (generation, research, review, governance), persistent memory, a governed task lifecycle with leases and fencing tokens, and a Next.js control center for real-time visibility. The portable core is published as the Metis Framework; the private operating repo layers identity, projects, and integrations on top.

Outcome

An agent workforce that compounds

→ A governed agent workforce that compounds across sessions instead of starting fresh.
→ Real-time control center with task boards, agent sessions, and typed decision queues.
→ Published the portable core as an open-source framework others can vendor into their own repos.