The security and control questions first — they're the ones that matter.
Every task runs in a fresh, isolated sandbox container hosted by Anthropic (the makers of Claude) — not on your machines and not on a shared box. The container is created for the task and destroyed after it. Your repo is mounted inside for the duration of the task only.
No. The engineering roles are instructed to always create a new agent/* branch and push that; you review and merge like any teammate’s PR. The QA role is read-only by design — it never commits. You stay the only one who can touch main.
You create a fine-grained token scoped to only the repos you connect, with Contents permission. We verify it against GitHub when you save it, store it encrypted (AES-256-GCM), and it is never placed inside the agent’s sandbox — git authentication happens at a proxy outside the container, so even the agent’s own code can’t read it.
Anthropic’s Claude family. By default the frontend and backend engineers run Claude Opus 4.8 (the most capable) and the QA, DevOps, marketing and UI/UX roles run Claude Sonnet 5. You can change the model per role in the dashboard — cheaper models for lighter roles is a common setup.
Every task’s exact token usage is metered and priced, and you see per-task cost in your dashboard. You set a hard monthly budget per server: once it’s reached, the team politely refuses new tasks until you raise it or the month rolls over. A typical small task costs $0.10–$0.50.
Three things: a Discord server you manage, a GitHub repo, and about five minutes. Sign in with Discord, invite the bot, paste a repo token, and mention your first agent. No credit card needed during the trial.
Yes — each role’s system prompt (its job description), model, and enabled state are fully editable, and you can add custom roles like "backend", "mobile", or "docs". Changes apply from the role’s next task.
Recurring tasks on a cron schedule that report to a channel you pick — for example "every weekday at 7:00, run the test suite on main and summarize what changed" or a morning production health check. You manage them in the dashboard.
Something else on your mind? Contact us — a human answers.