Introducing factotem
A Mac mini in your office hires agents the way you hire staff — one on WhatsApp, one on email, one running scheduled work overnight. Your hardware, your data, your model of choice. Built by the team behind Kanban Pro.
Paradigm
Most AI assistants live on someone else's infrastructure — your agents, your data, your control surface, all rented from a vendor. Factotem inverts that: the entire workforce runs on a Mac mini (or any server) you own, behind your network boundary, against credentials you hold. The codebase is small enough to read end-to-end, and the operator — you, at the keyboard — is always the highest-trust party in the system.
How it works
WhatsApp, Telegram, Slack, Discord, Gmail. Channels self-register at startup, and adding a new one is a skill — run /add-telegram inside Claude Code and you're connected. Trigger word configurable; messages flow through a per-group queue with global concurrency control.
Every chat group runs its agent in its own Linux container, with its own mount allowlist, its own CLAUDE.md memory, and its own SQLite session record. Bash inside the agent is safe because it runs inside the container, not on your host — and one group's context never leaks to another.
A local dashboard at :7842 over Tailscale. A signed macOS menu-bar Doctor that probes the stack every five seconds and exposes a typed-confirm Repair Stack action. A twelve-step wizard that takes a fresh Mac to "agent alive on WhatsApp" in roughly thirty minutes.
Why now
Capable agents are everywhere — but they live behind someone else's API, on someone else's terms, with someone else's outages. Factotem gives the operator the surfaces needed to actually run a workforce at home or for a small organisation: a dashboard, a menu-bar Doctor, a cold-start wizard, and a small forkable codebase. Not a managed cloud. Not a config dialog. A stack you own.
Model-agnostic by design
Factotem already speaks the Anthropic API format, so today it runs against Ollama on-device, Together AI, Fireworks, or any endpoint that speaks the protocol — not just Claude. The trajectory is full model agnosticism: cloud APIs, on-device weights, your own fine-tunes, future protocols. The orchestrator should care about coordination and isolation, not about which LLM lives behind the call.
In the wild
One Mac mini in your office runs the orchestrator. One agent answers WhatsApp; another watches your inbox; a third runs scheduled work overnight. Each in its own container, all on hardware you own. Tickets cross-link to Kanban Pro; the Brain corpus syncs via Google Drive. Roughly thirty minutes from npx claw-setup to your first hire.
A shared Mac mini for a household, friend group, or small organisation. Each chat group hires its own agents, with its own memory and scoped permissions. The segment admin runs the deployment without depending on any platform aggregator — and the same isolation that keeps one chat group from reading another's memory keeps your community's data on its own hardware.
Autonomous employees, on the servers you own, working for you alone.