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. mux 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.

curl -fsSL https://raw.githubusercontent.com/vxcall/cmux/main/install.sh | sh

Then add .worktrees/ to your .gitignore:

echo '.worktrees/' >> .gitignore

mux new <your feature name> --claude   # creates worktree + branch, runs setup hook, launches Claude
mux new <your feature name> --codex    # same, but launches Codex
mux new <your feature name>            # worktree only, no agent launched

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

You're building a feature:

mux new feature-auth --claude        # agent starts working on auth

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

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

Merge the bugfix when it's done:

mux merge fix-payments --squash
mux rm fix-payments

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

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

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

When mux new creates a worktree, it runs .mux/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:

mux 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.

• Worktrees live under .worktrees/<branch>/ in the repo root
• Branch names are sanitized: feature/foo becomes feature-foo
• mux new is idempotent on the worktree — if it already exists, it skips creation and setup
• Use --claude or --codex with mux new/mux start to launch an agent; omitting the flag only enters the worktree
• mux merge and mux rm with no args detect the current worktree from $PWD
• Pure bash — just git and the Claude CLI

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

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

MIT