The most-starred developer tool of the agent era is not an agent. Worktrunk, a Rust CLI by the developer known on GitHub as max-sixty, added roughly 524 stars in the past 24 hours according to GitHub's trending page, and a Reddit thread this week put the project at about 9,300. Its pitch fits in one line: git worktrees, without the typing.
Worktrees let one repository be checked out into several directories at once, each on its own branch — precisely what running five coding agents in parallel requires, because two agents in the same directory overwrite each other's files. Git has shipped the capability for years; the interface never caught up. Creating one means typing the branch name three times: `git worktree add -b feat ../repo.feat`, then `cd`, then the agent command.
Worktrunk addresses worktrees by branch name and derives the path from a template. `wt switch feat` jumps in; `wt switch -c -x claude feat` creates the worktree and launches Claude Code inside it; `wt remove` tears down directory and branch. Around those verbs it wraps what teams otherwise script by hand: hooks on create, pre-merge and post-merge; commit messages generated from the diff; a merge command that squashes, rebases and fast-forwards in one step; an interactive picker with live diff previews; per-worktree dev-server ports via a `hash_port` template filter; and the one that matters at scale — copying build caches instead of reinstalling `node_modules` or rebuilding `target/` in every checkout.
It is MIT/Apache-2.0, installs via Homebrew, Cargo, pacman, conda or winget, and on Windows also ships as `git-wt` because `wt` collides with Windows Terminal. It is a single-maintainer project that calls itself "the most popular git worktree manager" — a claim its star graph supports and which nothing in the git ecosystem currently contests.
What the trend says is more interesting than the tool. The bottleneck of agentic coding has shifted from model capability to environment management: whose filesystem, whose branch, whose port, whose cache. GitButler shipped worktree support in its 0.23 release last week; Orca, a newer entrant, fans a single prompt across five agents in five worktrees and lets you merge the winner. Anthropic's own Claude Code guidance now recommends the worktree pattern by name.
None of that is AI. All of it is the scaffolding that makes running AI in parallel more than a demo — and the reason a one-person CLI can matter more to a team's throughput than the next model release.
Watch next: whether worktree orchestration gets absorbed into the agent products themselves (Anthropic, OpenAI and Google all ship their own CLIs) or stays a third-party layer. The long-term value of a tool like this sits entirely in the second outcome.
