San
A 12MB open runtime for agents that work together
8 followers
A 12MB open runtime for agents that work together
8 followers
San is an open-source ~12MB terminal runtime for AI agents: one binary, ~0.01s cold start, any model. Delegate to subagents in parallel or arrange them as workflows, and let sessions on one machine message each other in a group to review and correct each other's work. It learns from your sessions into memory and skills, runs from laptops to Raspberry Pi, Jetson and CI, and its agent loop ships as a Go library (genai-io/sdk-go).
Hi PH 👋
I kept rebuilding the same layer for every agent I made — the harness around the model. So we built one worth keeping and opened it up.
San is one ~12MB Go binary with ~0.01s cold start — small enough to run anywhere, from your laptop to a CI step, a container, or an edge device.
It's flexible by design: plug in any model and search backend, switch personas to keep each agent's context lean. It learns as you work, turning recent sessions into durable memory and reusable skills. And it plugs into what you already use — skills, subagents, plugins & MCP run unmodified.
It's an open, fast harness, made to be built on. Apache-2.0, and we'd love to build it out with you.
Tell us what you'd plug in 👇