Worktrunk: The Git Worktree Manager Built for AI Agents

Most git worktree managers get built to solve a human's problem: too many branches, too many terminal tabs. Worktrunk was built to solve an AI agent's problem instead. Its own GitHub description is blunt about it: a "CLI for Git worktree management, designed for parallel AI agent workflows."

It's a genuine git worktree manager for AI agents, not a wrapper with an AI feature bolted on. Right now it's sitting on GitHub's daily trending leaderboard with 10,028 stars, up 77 today — built almost entirely by one person.

What this git worktree manager for AI agents actually does

Strip away the AI framing and Worktrunk is a Rust CLI for git worktrees — the git feature that lets you check out multiple branches into separate directories at once, instead of stashing and switching inside a single working copy. Per the README, the core commands are wt switch (switch to or create a worktree by branch name), wt list (worktrees with live status), wt merge (squash, rebase, merge, and clean up in one step), and wt remove.

On top of that base, the README lists a feature set aimed at running several things in parallel rather than one thing carefully:

• Hooks that fire on worktree create, pre-merge, and post-merge.
• LLM-generated commit messages from diffs.
• An interactive picker with streaming CI status and diff/log/PR previews.
• wt switch pr:123 to jump straight to a pull request's branch.
• Shared build caches on APFS, btrfs, and XFS — the one that matters most for agent workflows, since ten worktrees can reuse target/ or node_modules/ without rebuilding or copying them.

It installs through Homebrew, Cargo, Winget, Arch's pacman, or Conda/Pixi. It's dual-licensed MIT or Apache-2.0, with no pricing page or paid tier anywhere in the repo.

The maintainer behind it

Worktrunk belongs to Maximilian Roos, GitHub handle max-sixty, based in Los Angeles. His profile doesn't list an employer — it lists projects instead: he's also the maintainer of PRQL (a "pipelined SQL replacement," ~10.9k stars), xarray (N-D labeled arrays for scientific Python, ~4.2k stars), and the Rust snapshot-testing library insta (~3k stars), among smaller tools. Worktrunk is the newest and, by star count, now the biggest thing on that list. A commit history filtered to his own username shows at least 35 commits authored directly by him in the ten days before this post — not just bot-filed dependency bumps.

A year (or so) of shipping fast

GitHub's own release history doesn't make it easy to pin an exact birthday. The releases page lists the current latest tag as v0.81.0, published October 20, 2026. Paging all the way back — 16 pages deep, with page 8 landing on February-dated entries — the oldest tag on record is v0.1.0, which GitHub's own release page says was simply "released this 03 Nov," with no year printed anywhere on that page.

The project's LICENSE file carries a copyright line reading "Copyright (c) 2025 Maximilian Roos," which is at least consistent with a late-2025 start. Whatever the precise date, going from tag 0.1.0 to 0.81.0 across roughly 150-plus tagged releases in well under a year is a fast, prolific release cadence for a single maintainer.

That speed shows up in the open-issue count too, in a less flattering way. GitHub numbers issues and pull requests from one shared sequence, and the issues tab currently shows 43 open issues and 24 open pull requests, with the newest items numbered around #4672–#4673 — meaning close to 4,700 issues and PRs have been filed against this repo in that same stretch. That's a heavy amount of inbound traffic for a single maintainer to triage, similar to the load we've seen on other fast-growing solo repos like ai-job-search.

The part that makes this an indie-builder story, not just a dev-tool story

Worktrunk doesn't just help you run Claude Code and Codex in parallel — it ships its own integrations for them, the same instinct behind other AI agent skills built directly on top of those tools, like Diagram Design. The repo root holds a .claude-plugin directory with a marketplace.json manifest, a separate .codex directory for Codex Cloud setup, and a gemini-extension.json file. Per that manifest, the Claude Code plugin provides "automatic activity tracking (🤖/💬 indicators in wt list showing active Claude sessions), and a /wt-diff pane showing each worktree's changes." The project's CHANGELOG shows this agent-integration layer still actively growing: v0.77.0 added the terminal-based agent Pi to the list of tracked sessions. The latest, v0.81.0, routes OpenCode's own worktrees through Worktrunk after its plugin is reinstalled.

The repo also governs itself the same way it expects agents to behave. Its AGENTS.md — the Worktrunk-specific equivalent of a CLAUDE.md file — lays out rules for any coding agent working on the project:

• Treat config formats and CLI arguments as "protected interfaces," and new commands, flags, and config keys as "product decisions" — both need care.
• Load specific named skills (writing-user-outputs, running-tend, release) before touching certain code.
• Never hand-edit CHANGELOG.md in a feature PR — release entries are auto-generated from commit history by a dedicated release skill.

And then there's the clearest evidence this isn't just a tool for agents but a tool built with them. Recent pull requests opened and merged directly by max-sixty carry closing lines that read, verbatim, "This was written by Claude Code on behalf of max-sixty" — repeated on a second PR. On a third, the same line names Codex instead: "This was written by Codex on behalf of max-sixty." All three merged between October 28 and November 1, 2026.

That's not a marketing claim about "built with AI." It's an attribution line left in the project's own, publicly readable pull request history — on a tool whose entire purpose is making it safer to run those same agents side by side.

The claims worth taking with a grain of salt

The README makes two claims that don't come with any data attached. It says Worktrunk "was released at the start of the year, and has quickly become the most popular git worktree manager" — a comparative claim with no benchmark, star-count comparison, or download numbers cited anywhere in the repo to back it up, and one that sits a little oddly next to that undated "03 Nov" earliest tag. It also says the project is "built with love (there's no slop!)" — a nice line, but also just the maintainer's own characterization of his own tool. Treat both as self-reported color, not independently verified fact.

What's still rough

The open-issue list is a useful counterweight to the trending-page star count. Current reports include a recurring Windows timeout in a fork-remote-URL test (#4672) and a "Bot temporarily unavailable" report (#4673), both filed the same week. Longer-standing feature requests — making wt list the default action (#4260) and adding sort support to wt list (#3922) — have sat open for months. None of these are dealbreakers, but they're the normal backlog of a fast-moving CLI tool maintained by one person fielding nearly 4,700 issues and PRs worth of inbound traffic.

What this generalizes to

Worktrunk isn't a SaaS with an MRR screenshot — there's no pricing page, no company, just a dual MIT/Apache-2.0 license and a Cargo crate. That's closer in spirit to how openGym or ai-job-search stay free than to how a funded startup operates.

What it demonstrates instead is a specific shape of indie-builder leverage that's new to this cycle: a single maintainer, already juggling several other open-source projects, shipping a dev tool fast enough and specific enough to AI-agent workflows that the tool's own commit history becomes proof of the workflow it's selling. If you're going to claim your project makes running Claude Code and Codex in parallel easier, leaving "written by Claude Code" in your own merged pull requests is a harder claim to fake than a README badge.