# What you will learn

• What agentbox actually does: parallel Claude Code and Codex agents, each in its own git worktree.
• How to install it and point it at your existing subscription plan.
• How the 'answer like email' and board views change the way you supervise agents.
• One small example: running two agents on two features without merge chaos.
• The real limitations — worktree hygiene, disk churn, and what it does not solve.

# What is agentbox, and why should you care?

If you have ever tried to run three Claude Code sessions at once, you already know the pain. Two agents edit the same file, one silently reverts the other's change, and you spend the next twenty minutes reconstructing what happened from terminal scrollback. savannahfeder/agentbox is an attempt to fix that with a desktop GUI on top of git worktrees: every agent gets its own checkout, so parallel work stops colliding.

The repo describes itself plainly: a desktop GUI for running dozens of Claude Code and Codex agents in parallel, each in its own git worktree, with two interaction modes — answer them like email, or use the board view — and the claim that it is free, open source, and runs on the plan you already pay for. That last clause matters more than it looks. It means agentbox is not selling tokens or a hosted orchestration layer; it is a supervisor for the subscription you already have.

It is also young. The project is written in JavaScript, surfaced on GitHub Trending recently, and the star count is in the low hundreds, not thousands. Treat it as a promising early tool, not infrastructure you bet a team on.

# How do you install it and get your first agents running?

Because agentbox is a desktop app, the install path is the usual Node-based one. Clone the repo, install dependencies, and launch the dev or packaged build. The exact scripts live in the repository's package.json; the pattern below is what you will typically run:

git clone https://github.com/savannahfeder/agentbox.git
cd agentbox
npm install
npm run dev

Before you do that, make sure the prerequisites are in place. You need Node and a working git install, and you need Claude Code and/or Codex already authenticated on your machine. agentbox is a front end for those CLIs — it does not replace their auth or their models. If your Claude Code session already works in a terminal, agentbox has something to drive.

Once the app is up, the mental model is simple. You point it at a repository, it creates worktrees for you, and each agent session lives inside one. When you launch an agent, you are not opening a chat window so much as opening a worker with its own branch and its own files.

# What does 'answer them like email' actually feel like?

This is the feature that makes agentbox more than a launcher. Instead of watching a live terminal stream, you get a queue of agent messages — status updates, questions, blocked states — and you respond in batches. It is asynchronous supervision. You check in, read what the agents did, reply, and move on.

That is a real shift in how you work. Terminal-based agent orchestration encourages you to sit and watch; an inbox encourages you to delegate and return. If you have read about how V7 gives AI agents institutional memory, you will recognize the same instinct: the bottleneck is not model capability, it is the human's attention bandwidth. agentbox attacks that bottleneck directly by making agent output something you triage rather than something you stare at.

The board view is the complement. Where the inbox is chronological, the board is spatial — agents as cards, statuses as columns. If you are running a dozen agents against a dozen features, the board is how you see at a glance which ones are stuck. Neither view is novel in isolation; the combination, tied to real worktrees, is the useful part.

# How do git worktrees keep parallel agents from stepping on each other?

A git worktree is a second working directory attached to the same repository, checked out to a different branch. Two agents in two worktrees can edit the same file paths without touching each other's copies. That is the whole trick, and it is a good one: it turns a merge-conflict problem into a merge-review problem.

Here is the small example worth trying first. Pick a repo with two independent pieces of work — say, a login bug and a settings-page refactor. Create two agents in agentbox, each in its own worktree, and give each a scoped task. Let them run. When both finish, you review the branches one at a time and merge them in whichever order makes sense. Compare that to running both agents in one checkout and discovering at the end that one overwrote the other's changes.

This is the same lesson that shows up across the agent-tooling conversation: the model is rarely the weak link. AI codes better than most devs — that was never the job makes the point that the job is coordination, review, and integration. agentbox is a coordination tool, and it should be judged on that basis.

# Where does agentbox fall short?

Several places, and you should know them before you invest a weekend.

• Worktree hygiene is on you. agentbox creates worktrees; it does not automatically prune them. Run enough agents and you will accumulate branches and disk usage that need cleanup. Budget time for git worktree list and git worktree remove.
• It is a supervisor, not a reviewer. The inbox tells you what agents say; it does not tell you whether their code is correct. You still need to read diffs.
• It is young. Low hundreds of stars, JavaScript, active but unproven. Expect rough edges in the GUI and in edge-case handling of unusual git setups (submodules, large monorepos).
• It depends on the CLIs. When Claude Code or Codex changes its output format, agentbox has to chase it. That is a structural risk of building on someone else's interface.

The comparison that matters is not agentbox versus a hypothetical polished product. It is agentbox versus your current habit of opening four terminal tabs and hoping. On that comparison, the worktree isolation alone is worth the install. For more on where agent orchestration is heading, see SwarmEvolve skips the chain of thought, keeps the swarm — the pattern of many small workers under a supervisor keeps recurring for a reason.

My advice: install it this week if you already run Claude Code daily. Start with two agents, not twenty. Learn the worktree cleanup rhythm before you scale up. The tool is early, but the idea — parallel agents with real filesystem isolation and an inbox to manage them — is the right shape for how coding agents will actually be used.