cmux — tmux for Claude Code

Run a fleet of Claude agents on the same repo — each in its own worktree, zero conflicts, one command each.

Claude Code works best when it has full ownership of the working directory. Want two agents working in parallel? They'll stomp on each other — conflicting edits, dirty state, broken builds. You need separate checkouts.

Git worktrees are the perfect primitive for this. They share the same .git database but give each agent its own directory tree — no cloning, no syncing, branches stay in lockstep. cmux wraps the entire worktree lifecycle into single commands so you can spin agents up, jump between them, and tear them down without thinking about it.

You wanna go fast without losing your goddamn mind. This is how.

Install

curl -fsSL https://github.com/craigsc/cmux/releases/latest/download/install.sh | sh

Then add .worktrees/ to your .gitignore:

echo '.worktrees/' >> .gitignore

Quick start

cmux new <your feature name>       # creates worktree + branch, runs setup hook, opens Claude

That's it. One command, one agent, fully isolated. See Workflow for the full loop.

Commands

Command	What it does
`cmux new <branch>`	Create new worktree + branch, run setup hook, launch Claude
`cmux start <branch>`	Continue where you left off in an existing worktree
`cmux cd [branch]`	cd into a worktree (no args = repo root)
`cmux ls`	List active worktrees
`cmux merge [branch] [--squash]`	Merge worktree branch into your primary checkout (no args = current worktree)
`cmux rm [branch \| --all]`	Remove a worktree and its branch (no args = current, `--all` = every worktree with confirmation)
`cmux init [--replace]`	Generate `.cmux/setup` hook using Claude (`--replace` to regenerate)
`cmux update`	Update cmux to the latest version
`cmux version`	Show current version

Workflow

You're building a feature:

cmux new feature-auth        # agent starts working on auth

Bug comes in. No problem — spin up another agent without leaving the first one:

cmux new fix-payments        # second agent, isolated worktree, independent session

Merge the bugfix when it's done:

cmux merge fix-payments --squash
cmux rm fix-payments

Come back tomorrow and pick up the feature work right where you left off:

cmux start feature-auth      # picks up right where you left off

The key distinction: new = new worktree, new session. start = existing worktree, continuing session.

Setup hook

When cmux new creates a worktree, it runs .cmux/setup if one exists. This handles project-specific init — symlinking secrets, installing deps, running codegen. If no setup hook exists, you'll be prompted to generate one.

The easy way — let Claude write it for you:

cmux init

Or create one manually:

#!/bin/bash
REPO_ROOT="$(git rev-parse --git-common-dir | xargs dirname)"
ln -sf "$REPO_ROOT/.env" .env
npm ci

See examples/ for more.

How it works

Worktrees live under .worktrees/<branch>/ in the repo root
Branch names are sanitized: feature/foo becomes feature-foo
cmux new is idempotent on the worktree — if it already exists, it skips creation and setup, but still launches a new Claude session
cmux merge and cmux rm with no args detect the current worktree from $PWD
Pure bash — just git and the Claude CLI

Tab completion

You never have to remember branch names. Built-in completion for bash and zsh — automatically registered when you source cmux.sh, no extra setup.

cmux <TAB> — subcommands
cmux start <TAB> — existing worktree branches
cmux cd <TAB> — existing worktree branches
cmux rm <TAB> — worktree branches + --all
cmux merge <TAB> — worktree branches
cmux init <TAB> — --replace

License

MIT