What changed
- Approve the team before it starts
- Follow results without losing the conversation
- Compact rows keep the task name, status and actions aligned. Results and team details expand when needed; file comparisons no longer fill the conversation with large cards.
Approve the team before it starts
Enable Agent team in task settings. The parent chooses from ready, compatible connections and explains each assignment. Review the prompt, agent and model; select the helpers you want to start.
Four concurrent helpers by default, up to sixteen in settings. Each works in a separate copy. Helpers cannot create more teams; repeated starts are bounded.
Follow results without losing the conversation
Child conversations can open in the right-hand panes. The parent can read their results, send clarifications and request pause or stop. Closing a pane does not stop its task, and changes are never merged automatically.
Compact rows keep the task name, status and actions aligned. Results and team details expand when needed; file comparisons no longer fill the conversation with large cards.
Choose a supported connection
Teams support isolated Claude Code, Codex and Z.ai Coding Plan connections, plus scoped OpenRouter and OpenAI-compatible model connections on the same computer. CLI agents without verified isolation, Cloud, direct-folder tasks, Dangerous mode and shared task budgets are not supported in teams yet.
Dangerous mode has a shorter confirmation: acknowledge the risks and enable it. The outside-project access warning remains explicit; plan review and questions still need your answer.
Web, Linux and Windows. macOS builds are released separately.
Reactions are public totals. Your choice is saved in this browser.