agent-manager

agent-manager is a Golang TUI for running several coding-agent CLIs at once without hunting through terminal tabs. Each session is your own installed CLI running unmodified in a persistent tmux session, so logins, subscriptions, config files and MCP servers behave exactly as they do in a plain terminal. The manager adds one list of all sessions with live status, grouped into a foldable project tree, and a way to act on any of them without attaching. The README's tagline is "the fastest workflow for every AI coding agent", which is marketing rather than a measured claim.

How it works

Sessions live on a private tmux server named agentmgr, in an am_* namespace, so they never mix with your own tmux and survive the manager quitting. From a plain shell you can still reach one:

tmux -L agentmgr ls
tmux -L agentmgr attach -t am_<id>

Status detection supports Claude Code, OpenCode, Codex, Grok Build, Gemini CLI, Pi, Command Code and Hermes Agent. The launch, revive, fork and status rules for each tool are compiled into the binary, so upgrading agent-manager updates them all at once. The flip side is that when one of those CLIs changes its interface, reading it correctly again waits on an agent-manager release.

A session can start in its own git worktree (<repo>-worktrees/<name> on branch am/<name>), toggled per session or by default. Every MCP-capable session is launched with tools that let one agent spawn another, send it a message and wait for it to finish.

Using it

space is the central key: it sends a prompt straight into the selected session's pane, or spawns a new agent when a group is selected. n creates a session with a name, tool, directory, optional first prompt and group. x kills a session to free memory and v revives it on its own conversation; R restarts it on an empty context, and f forks the conversation into a separately named session. T opens a shell under the selected agent or group for builds and git.

The review screen is the distinctive part. ctrl+r opens full-file, syntax-highlighted diffs of what an agent changed; c leaves a comment on a line, and C sends all comments back to the agent's pane as one numbered review round. Sent comments stay visible as open or handled. That is the review loop of reviewing-ai-code made cheap enough to run per session, which does nothing about its throughput limit but removes the copy-paste.

brew install yoanwai/tap/agent-manager
curl -fsSL https://raw.githubusercontent.com/YoanWai/agent-manager/main/install.sh | sh

Runs on macOS and Linux, and on Windows only inside WSL2. Needs tmux 3.1 or newer and git; the install script checks checksums and offers to install missing dependencies. The end-to-end tests drive a real tmux server. Cost tracking and mouse-driven navigation are listed as not there yet.

Neighbours

showagent works on finished sessions on disk (browse, resume, convert between agents), where agent-manager supervises live ones. kanbots gives each task a worktree and a card on a desktop board; agent-manager does the same in the terminal with a tree instead of columns. fresh-editor builds the same per-worktree agent workspaces into an editor, and agentsview covers the cost analytics agent-manager lacks. It runs agents on the host with your credentials and no isolation, unlike superhq.

On watchlist: two months old, single author, heavy launch promotion (Trendshift, Product Hunt, Peerlist badges), and status detection that has to keep pace with eight CLIs that change often.

Apache-2.0, 432★, 40 forks, created 2026-07-15, pushed to on the day of ingest. Repo: https://github.com/YoanWai/agent-manager