Overseer — Personal AI ops layer

Multi-agent runtime I built and run across a VPS and a Mac. Ten production services coordinating over Postgres LISTEN/NOTIFY: a brain that decides, a gateway on :8768, a remote MCP at mcp.sentryaithermal.com that Claude and ChatGPT connectors call into, a commitment promoter, an action executor gated by one-tap approval. ~17,500 durable memories in a vector store. ~65 proposed actions a week.

Three things it does without me asking

• Commitments I make in iMessage or voice become tracked tasks with deadlines; the system chases me at the deadline.
• Before any “meeting in 20 minutes” calendar push, it scans iMessage, mail, and recent voice transcripts for a sick-day or reschedule signal — caught a stale meeting on a colleague sick-day last week.
• Every Claude and Codex session writes durable rules to the same store, triaged daily by a smaller model, auto-committed to GitHub every 30 minutes. Sessions on a new host inherit context without me retyping.

How it is built

Postgres for shared state and event bus. FastAPI gateway. A Next.js operator console at cp.jcornelius.net. A remote MCP server that Claude Desktop and ChatGPT custom connectors authenticate against. Approvals are one-tap from Pushover, the operator console, or a connector call. Nothing in the proposed-action queue executes without that gate.