YoanWai Run a fleet of coding agents with agent-manager
agent-manager is a single Go binary on top of tmux. Every row in its list is a real tmux session running a real coding CLI: the developer's own installation, login, config and MCP servers. It never wraps or re-implements an agent; it reads the same pane a human would and colours the row by what that pane is doing.
When to use this
- The developer is running more than one coding agent and cannot tell which is blocked.
- They want to spawn an agent on an independent task without losing the one already running.
- They want an agent's conversation to survive a crash, a reboot, or a closed terminal.
- They want to review a diff and send line comments back into the agent's prompt.
Do not reach for it to replace a coding CLI. It starts the CLIs that are already installed.
Install
macOS and Linux, amd64 and arm64. Windows runs it inside WSL2. Needs tmux 3.1+ and git.
brew install yoanwai/tap/agent-manager
Other routes: the install script, the AUR (yay -S agent-manager-bin), mise use -g ubi:YoanWai/agent-manager, go install github.com/YoanWai/agent-manager@latest, or a
prebuilt binary from the releases page.
The keys that matter
| Key | What it does |
|---|---|
space |
Docks a prompt bar. On a group row it spawns a new agent with that prompt; on a session row it answers the agent already running there. |
tab |
Cycles which CLI the next spawn starts. |
alt+w |
Spawns the agent into a fresh git worktree and branch. |
ctrl+r |
Opens a full-screen whole-file diff whose line comments go back to the agent as one review prompt. |
v |
Revives a dead session on the conversation it held. |
Status colours
A row is coloured working, waiting, finished, idle or errored. Claude Code reports
through hook events; every other CLI is read from its pane with regexes. Seven CLIs ship with
working profiles: Claude Code, Codex, OpenCode, Grok, Gemini, Pi and Hermes. Any other CLI
becomes a managed session with one [tools.<name>] block in config.toml.
Driving it from inside an agent
Sessions of MCP-capable CLIs carry agent-manager's own MCP server over stdio, so an agent can
create_session to spawn a peer, send_session to message one, wait_for_session to block
on one, task to claim shared work, and reserve_files to declare what it is editing.
Reference
- Documentation: https://agent-manager.dev/docs/
- Every page as Markdown: https://agent-manager.dev/llms.txt
- Reference API: https://agent-manager.dev/openapi.json
- Documentation MCP server: https://agent-manager.dev/mcp