os

os is a persistent personal agent that I build and use every day. It runs on my own server with a web UI and a desktop client, and it has files, terminals, browser panes, and tools connected over MCP and OpenAPI: mail, calendar, finances, Home Assistant, and my Git forge.

Problem

Chat assistants forget everything and can’t safely act on my systems. I wanted one agent that keeps context over weeks, runs routines on a schedule, and asks before doing anything risky.

Work

• Routines. Scheduled jobs such as mail triage, a morning brief, and transaction categorization, each defined as a reusable skill.
• Approval-gated actions. Writes that matter ask for approval, naming the exact target and effect.
• Long-running jobs and subagents that survive restarts, with the results recorded.
• Deployed from platform like any other service, with separate prod and dev channels.

Stack

TypeScript, a Node server, web and Electron clients, MCP, Nix.

Status

Active and in daily use. Source releases are published on my forge.