Over the last months both my fleet of agents grew rapidly and I was getting more and more context from different sources, including videos, calls, chats. Feeding the inbound context into agents became more and more time-consuming, so together with my team we've built a tool to gather the context live and store it in an accessible manner for both humans and agents - so we can stop copy pasting things to the terminal.

A summary of what is ready: - For now MacOS only, other platforms support is in development - Knowledge graph accessible via MCP, web UI, and is shareable - Current benchmark is 82.8% IPR on LongMemEval (ICLR 2025) - HUD overlay - ambient agent wrapper advising on your context, plugs in CC, Codex, Goose, Junie - Default mode needs an OpenRouter API key (token usage is tracked); paranoid mode relies on local models (with a quality trade-off) - An architectural choice we are proud of: knowledge sharing between users is peer-to-peer, over webRTC. For now it does require a call to a redirector server (which does not see the data), but we are planning on getting rid of it all together. The recipient's agents pick up the context via MCP.

Sinain is MIT-licensed.

Site: https://anthillnet.com Demo: https://youtu.be/EWLcVtPPbac Repo: https://github.com/anthillnet/sinain-hud Install: npx @geravant/sinain@latest start

I would be glad to answer anything in the comments.