An AI agent that runs itself.

Claude, GPT, Gemini, Groq, local models — routed per task, swappable, never locked in.

Run it fully offline with a local model (Ollama) — telemetry-free, no API keys, nothing leaves your machine.

Short-term context, durable long-term notes and project files — it remembers what matters.

Curates its persistent memory across sessions, distills lessons from past work, and self-diagnoses its own failures to recover — so it gets sharper the more you use it.

Composable skills it loads on demand: research, files, deploys, messaging, and your own.

Spawns isolated workers for long or parallel jobs that survive a parent abort — on Claude or any provider — and reports back when they finish. An agent runs with the same rights as the chat it came from, so it never stalls in front of a prompt nobody is there to answer.

Run several separate conversations at once — one client or project per lane, each with its own memory, persona, folder and daily budget cap. Lanes of one workspace share what they learn only when you opt in; isolation is the default.

Nudge a running task mid-stream with /btw — it adapts on the fly, no restart — or stop it instantly.

Speaks and listens — voice notes in, spoken answers out, in multiple languages.

Exec allowlist, DM pairing and locked-down secrets out of the box — and everything it reads (email, web, documents, MCP connectors, uploaded files) is fenced against prompt injection. One switch lifts the command and write guards completely if that is what you want; responsibility is then yours. Who may use the bot at all, and the fence around outside content, stay untouched either way. Documented threat model, audit CLI included.

Preflight checks, heartbeat monitoring, crash forensics and a self-healing supervisor that escalates to a controlled self-restart when needed — it recovers before you notice. Should the main loop stall anyway, a second thread is watching it and records what it is stuck on while it is still stuck.

Scheduled jobs and recurring check-ins — it acts on a clock, not only on a prompt. Search the job list and filter it by active, paused, failing, type, model or origin; pause, resume or delete several jobs in one go.

The same bot in a browser: every lane side by side, a cron page, workspaces you can create and remove, settings, a system doctor — and short synthetic cues for sent, finished, agent started, error, provider switch, each one switchable on its own. Works on a phone as well as at a desk.

When a provider hits its quota, the bot recognises it from the limit event, the error or the stated reset time, locks that one account until then, and answers on the next link of your chain. You get one line about it — once, not on every message.

Shell and Python run for up to ten minutes, a single call for up to thirty if it asks — enough for a build, an npm install or a data run, not just an ls. Files are read asynchronously, so one blocking access never halts the whole bot.