Worktrunk: Git Worktree CLI for Parallel AI Agents

Worktrunk is a Rust-based CLI that creates, lists, merges, and removes multiple Git worktrees with short commands. It is purpose-built for running AI coding agents such as Claude Code and Codex in parallel within a single repository. It is developed by max-sixty and published at max-sixty/worktrunk on GitHub. As of the time of writing (September 22, 2026), it has 8,320 stars and 290 forks, is dual-licensed under MIT OR Apache-2.0, and its latest release is v0.79.0 (published September 21, 2026).

In short, worktrunk fixes the clunky UX of Git's native worktree feature. Even starting a single new worktree with plain git requires typing the branch name three times: git worktree add -b feat ../repo.feat, then cd ../repo.feat. Worktrunk compresses this into a single wt switch -c feat, and the -x flag can launch an agent in the same step. As AI agents become able to handle longer tasks unsupervised, managing 5-10+ parallel working directories in one repository has become practical, and worktrunk is built specifically to support that workflow.

The command name is wt, but on Windows it conflicts with Windows Terminal's own wt command, so the Winget install provides it under the alias git-wt instead.

Worktrunk is available via Homebrew, Cargo, Winget, Arch Linux's pacman, and Conda/Pixi. The wt config shell install step after installation sets up shell integration, which is what lets worktrunk's commands change the working directory of the current shell session (without it, directory changes stay confined to a subshell).

Worktrunk is free, open-source software; no paid tier is officially offered as of the time of writing.

The shortest workflow is to create a worktree for a feature branch, check status with wt list, and merge with wt merge when done.

Once work is done, there are two paths: the PR workflow, where you commit, push, and open a PR manually and clean up with wt remove after it merges; or the local merge, where wt merge generates a commit message from the diff and handles rebase, merge, and removing the now-unneeded worktree automatically.

To run AI agents in parallel, the -x flag launches an agent right when the worktree is created.

Combined with the -x flag, a post-start hook that installs dependencies or starts a dev server means a single wt switch call leaves an agent ready to work immediately. In the Claude Code integration, a /worktrunk skill assists with configuring things like LLM commit messages, project hooks, worktree path templates, and shell-integration fixes, and agents started with isolation: "worktree" create worktrees following worktrunk's naming conventions, hooks, and lifecycle management. A statusline is also available via wt list statusline --format=claude-code, showing model name and context usage in the editor.

Tools for running parallel AI agents broadly split into CLIs that only manage worktrees and GUIs that manage the whole parallel-agent workflow. Worktrunk is a leading example of the former; a leading example of the latter is Orca, the open-source agent development environment covered in an earlier article on this site.

In short, worktrunk stays narrowly focused on creating, viewing, and cleaning up worktrees, and can serve as a foundation underneath a full environment like Orca or a multiplexer like herdr. The official README itself shows an example pairing worktrunk with Zellij, a terminal multiplexer, across tabs.

- Best suited to people comfortable working from the command line. If you want to see worktree and agent status visually, Orca or Conductor fit better
- A single lightweight Rust binary with no Electron dependency, making it easy to embed in CI or run over SSH
- Directory changes from wt switch won't apply to the current shell unless wt config shell install has been run
- On Windows, the wt command conflicts with Windows Terminal, so it installs as git-wt; forgetting the alias breaks invocations
- Shared build caches rely on copy-on-write support in APFS, btrfs, or XFS and have no effect on unsupported filesystems
- Auto-generated commit messages and AI summaries in wt merge involve LLM calls, so API costs or rate limits may apply (check official sources for current pricing details)
- As a project maintained by a single independent developer (max-sixty), organizations should apply their normal OSS vetting process before adopting it in a company setting

Worktrunk simplifies the foundational part of running parallel AI agents on Git worktrees — creating, listing, merging, and removing them — through a focused CLI. For a fully integrated GUI, Orca fits better; for terminal-based multiplexing of many agents, herdr; for heavier diff review, Hunk; and for shared desktop GUI control across people, Mosaic. In practice, a parallel-agent stack is best assembled by picking the right tool for each part of the workflow. We also support software development and AI consulting engagements for teams building out this kind of setup — reach out via contact if that would help.