bise

You talk to 'main', a team lead agent that stays always listening, knows the whole repo, splits your ideas into jobs, and runs the other agents. Peek into any agent when you want, but no tabs or sessions to manage.

Agents message each other before they collide, hand work off, create git worktrees when they need their own copy of the repo (and clean them up afterward), and each commits its own work.

Main answers agents' routine questions the way you would and tells you why; only real decisions reach your inbox. You can interject, add tasks, or change direction mid-run without stopping jobs.

Bring your own API key (Anthropic, OpenAI, Mistral and more), assign a different model and reasoning level per agent, connect unlimited MCP servers, and load Agent Plugins (skills, MCP servers, hooks) from ~/.agents/plugins or your repo. Tools are called through code so many servers don't fill the c

Press ctrl+r twice to talk to main out loud (/voice picks model, voice, and language). Agents can open pages in their own background tab group with your logins to read, click, fill forms, and take screenshots (off until /computer-use). Paste or drag images into the thread as chips.

In repos that use pull requests, every change gets its own branch and PR and flows through your usual CI, review bots, and teammates, with a panel showing each PR's status. ctrl+` opens a real terminal in your repo that keeps running while hidden.

Restart or update mid-work (bise update) and agents resume immediately with no lost work; the thread, draft, and queue come back. Agent communication is formally verified with Bend: no message lost, none delivered twice, restarts preserve state, and nothing waits forever.

Main doesn't carry your code — agents read it with fresh contexts while main keeps the plan and reports, so the thread stays light. Long threads are compacted automatically, and any agent can read the full history of messages and tool calls, so no separate memory system is needed.