The git worktrees your coding agents leave behind, and how to clean them up

Multiple coding agents, such as Claude Code, Cursor, and Codex, operate within git worktrees, allowing simultaneous work. However, these worktrees can accumulate, consuming considerable disk space. A Claude Code user discovered 27 worktrees consuming 80 GB, while 61 Codex worktrees for a Rails app occupied about 100 GB (issue 63087, 35383).

These worktrees reside in specific directories: .claude/worktrees/ within each repository (earlier versions utilized ~/.claude-worktrees), ~/.cursor/worktrees/ for Cursor, and ~/.codex/worktrees/ for Codex, operating on detached HEADs rather than branches.

To inspect the space they occupy, use git worktree list in a repository and du -sh .claude/worktrees/* 2 /dev/null | sort -h. For Cursor and Codex, apply du -sh ~/.cursor/worktrees/*/* ~/.codex/worktrees/* 2 /dev/null | sort -h.

Before removing a worktree, verify its status with git -C path status --short. If no uncommitted changes exist, use git -C path status --short --ignored to check for ignored files like .env. For squash merges, confirm their absence by checking if the branch has been merged into main using gh pr list --head branch --state merged --json headRefOid and comparing it with git rev-parse branch.

To prevent the buildup of worktrees, adjust the settings: Claude Code allows lowering cleanupPeriodDays (default 30), Cursor enables lowering cursor.worktreeMaxCount (default 25), and Codex can reduce the limit of 15 kept worktrees in its settings. For a practical solution, Mopup, a Mac application, automates these checks and only removes finished worktrees, never using --force, ensuring no committed data is lost. Mopup is free to scan repositories.