Agent Manager
GitHub
A new CLI tool called Agent Manager lets you run Claude Code, Codex, OpenCode, Grok, and Gemini CLI side by side in tmux, all from one dashboard. No more juggling terminal tabs to see which agent finished and which one's stuck waiting on you.
Based on reporting by GitHub — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anyone running more than one AI coding agent at once knows the drill: five terminal tabs, one of them silently waiting on a permission prompt for the last twenty minutes, and no easy way to tell which is which without clicking through all of them. Agent Manager, a new open-source tool from developer YoanWai, is built specifically to kill that friction.
The idea is simple. Each agent — Claude Code, Codex, OpenCode, Grok Build, or Gemini CLI — runs in its own tmux session, so the work keeps going even if you close the manager itself. Instead of hunting through tabs, every session lands in a single list with live status, organized into a project tree you can fold, reorder, or rename. You can answer an agent without ever attaching to its session: hit space and a prompt goes straight into that session's pane, or spawns a fresh agent in whichever group you've selected. A session that's died can be brought back with its old id and history intact using v.
The review workflow is where the tool gets more opinionated. Pressing ctrl+r opens a full-file diff of whatever an agent just changed, syntax-highlighted and scrollable, with changed lines tinted so you're reading in full context rather than a stripped-down patch view. Leave comments on individual lines, then press C to bundle every note into a single review prompt that lands back in the agent's pane — the agent starts addressing feedback while the diff keeps refreshing in front of you.
Agent Manager also lets agents self-identify. Sessions can rename themselves via a shell command once they've figured out what they're actually working on, and they can declare which git repo and which branch they're diffing against, so review doesn't have to guess when a working directory holds multiple repos. MCP-capable agents get these commands exposed as native tools automatically, no manual setup required.
What it doesn't do yet is just as telling: no worktree creation, no cost tracking, no mouse navigation, and no way for agents to talk to each other. Installation is straightforward across platforms — Homebrew, a curl script, Go install, an AUR package, or prebuilt binaries — and on Windows it runs through WSL2, since tmux itself is a Linux and macOS creature.
My take — AI-written commentary, not fact-checked reporting
Multi-agent coding setups have quietly become normal, and the tooling around them is finally catching up instead of forcing everyone to duct-tape terminal tabs together. Building this on tmux instead of some bespoke daemon is the right instinct — it's boring infrastructure that already works and survives crashes. The missing cost tracking is the real gap here, though: once you're running five CLI agents in parallel, the API bill is exactly the kind of thing you want visible without doing mental math across five different vendors.
Read more about this at: GitHub