Because everything is tmux, a supervising agent can start a worker, send it instructions, interrupt it with a control character and read exactly what a human would see — which makes multi-agent work debuggable in a way message-passing frameworks usually are not. The parallel path is the strongest feature and the one with real prerequisites: each worker runs in its own Git worktree, so it only works inside a repository, and automatic merging is off by default for good reason.
A control layer over AI agents running in tmux panes. Each pane is addressed as session:window.pane, so one agent can start Claude Code or Gemini in another pane, talk to it, and see what is on its screen. Documentation is in Japanese.
- start_agent launches claude or gemini in a named tmux target, with extra arguments if needed
- get_agent_status reports one pane, a wildcard across a session, or everything, distinguishing running, authenticating, actively executing and stopped for each agent type
- send_message writes to a pane, including control characters such as C-c to interrupt and C-l to clear, with sendEnter controlling whether the message is submitted
- capture_screen returns a pane's contents, all of it or the last N lines — which is how a supervising agent reads what a worker actually produced
- parallel_implement runs the same task across several workers, each in its own Git worktree, with worker count, complexity and optional automatic merge
- get_parallel_status reports progress across parallel sessions: each worker's state, completion and errors
Tmux, with a session created before anything starts (tmux new-session -d -s multiagent), and the agent CLIs you intend to launch installed and authenticated. It runs as npx agent-collaboration-mcp. parallel_implement must run inside a Git repository, since each worker gets its own worktree. Automatic authentication of newly started agents is optional and needs Playwright MCP available to the same client; without it, a new agent's login is done by hand.
One command — npx agent-collaboration-mcp