Where coding agents keep their data on a Mac
A folder-by-folder map of Claude Code, Codex, Cursor, OpenCode and Orca: what each folder holds and whether to clean it.
7 min read · Updated October 9, 2026
Coding agents keep their history, checkpoints, caches and worktrees in your home folder, and each one does it differently. This page is a map: for each agent, the folders, what's in them, and what to do with them. "Own command" means the agent has a command for it, and that's the way to clean it. "Never" means deleting it costs you an account, settings or a database.
To see how much each folder takes, du only reads:
du -sh ~/.claude ~/.codex ~/.cursor ~/.local/share/opencode ~/orca 2>/dev/null
Claude Code
Claude Code keeps almost everything in ~/.claude, or in CLAUDE_CONFIG_DIR if you set it. It deletes most session files itself once they're older than cleanupPeriodDays, 30 days by default.
| Folder | What it holds | Clean it? |
|---|---|---|
/private/tmp/claude-<uid> |
Session scratchpads and pasted images | Yes, for sessions that have ended |
~/Library/Caches/claude-cli-nodejs |
MCP and error logs | Yes, logs only |
~/.claude/projects |
Session transcripts and each project's auto memory | Own command: claude purge |
~/.claude/file-history |
Checkpoint snapshots | Claude Code deletes old ones |
~/.claude/debug, shell-snapshots, paste-cache, image-cache |
Debug logs, shell setup, pastes, legacy images | Claude Code deletes old ones |
~/.claude/plugins |
Installed plugins and marketplaces | Never; use /plugin |
~/.claude.json, ~/.claude/settings.json, ~/.claude/.credentials.json |
Account and global state, settings, login | Never |
~/.claude/history.jsonl, ~/.claude/agent-memory |
Prompt history, subagent memory | Never |
Worktrees go to .claude/worktrees inside each repository. claude purge <path> --dry-run shows what purging a project would delete, including its auto memory. More in How to free up disk space taken by Claude Code and on the Claude Code page.
Codex
Codex keeps its state under CODEX_HOME, which is ~/.codex unless you change it. The CLI, the desktop app and the IDE extension share it.
| Folder | What it holds | Clean it? |
|---|---|---|
~/.codex/.tmp |
Leftover temporary git repositories | After a look; what they're for isn't documented |
~/.codex/cache, ~/.codex/log, ~/Library/Logs/com.openai.codex |
Cache and logs | Yes |
~/.codex/sessions, ~/.codex/archived_sessions |
Session logs | Own command: codex delete <id> |
~/.codex/worktrees |
Worktrees made by the app | Codex keeps the newest 15 and snapshots before deleting |
~/.codex/shell_snapshots |
Shell setup snapshots | Leave to Codex |
~/.codex/plugins/cache |
Plugins | Never; Codex manages them |
~/Library/Application Support/Codex, ~/.codex/sqlite, ~/.codex/*.sqlite |
App state and databases | Never |
~/.codex/auth.json, config.toml, history.jsonl |
Login, settings, prompt history | Never |
Deleting session files by hand leaves stale rows in Codex's database. codex delete takes a session id or name and deletes it permanently, so check which one first. More on the Codex page.
Cursor
Cursor splits its data between ~/.cursor, or CURSOR_CONFIG_DIR, and ~/Library/Application Support/Cursor.
| Folder | What it holds | Clean it? |
|---|---|---|
~/.cursor/projects |
Agent transcripts left behind when chats are deleted | Look first |
~/Library/Application Support/Cursor/CachedData, logs |
A cache folder per version, and logs | Look first; keep the newest versions |
~/Library/Application Support/Cursor/User/globalStorage/state.vscdb |
Chat and agent database, with session keys | Never by hand; only from inside Cursor |
~/.cursor/chats |
CLI chats, kept only on this Mac | Only from the Cursor CLI |
~/.cursor/worktrees |
Agent worktrees | Cursor keeps up to 25 per machine |
~/.local/share/cursor-agent/versions |
CLI versions | Leave to the updater |
~/.cursor/mcp.json, cli-config.json, hooks.json |
MCP, CLI and hook settings | Never |
Never edit state.vscdb while Cursor is open. More on the Cursor page.
OpenCode
OpenCode keeps its data in ~/.local/share/opencode and its cache in ~/.cache/opencode, or under XDG_DATA_HOME and XDG_CACHE_HOME if you set them. Since version 1.2, sessions are in a SQLite database.
| Folder | What it holds | Clean it? |
|---|---|---|
~/.local/share/opencode/snapshot |
Undo snapshots, kept as shadow git repositories | After a look; some outlive their projects |
~/.local/share/opencode/storage |
Session files from before 1.2, copied into the database | After a look |
~/.local/share/opencode/log, ~/.cache/opencode |
Logs, and the provider package cache | Yes; OpenCode rebuilds the cache |
~/.local/share/opencode/opencode.db |
Session database | Own command: opencode session delete <id> |
~/.local/share/opencode/worktree |
OpenCode worktrees | From the OpenCode app |
~/.local/share/opencode/tool-output |
Tool output | Leave to OpenCode |
auth.json, mcp-auth.json |
Credentials | Never |
opencode session list shows the session ids. More on the OpenCode page.
Orca
Orca runs agents in git worktrees, so its data is your code.
| Folder | What it holds | Clean it? |
|---|---|---|
~/orca/workspaces/<repo>/<name> |
Workspaces, each a git worktree | One by one, with the checks for any worktree |
.orca-worktree-trash in a repository |
Workspaces Orca already removed | After a look |
~/orca/workspaces/.orca-preparing |
Setups that stopped halfway | After a look |
Treat each workspace like any worktree: check for uncommitted work, ignored files and unpushed commits, then remove it with git. See How to remove git worktrees left by coding agents, safely and the Orca page.
The rule of thumb
- Logs, caches and scratch space are the safe part. They come back or don't matter.
- Session history has an owner. Use the agent's own command, or the agent's database ends up out of step with its files.
- Accounts, settings, credentials and databases stay. Deleting them logs you out, resets your setup or breaks the agent.
- Worktrees hold code. Check them with git before removing anything.
Doing it with Devbroom
Devbroom knows these folders and treats each one the way its agent does. What the agent cleans itself it leaves to the agent, what has its own command it does or shows with that command, and accounts, settings and databases are never touched. Every item gets a verdict, Ready, Review, Blocked or Unknown, with its reasons. Cursor's data is only read for now, and so is OpenCode's if its database format doesn't match what Devbroom expects.
What you confirm moves to the Trash, and Put back returns it from History. The full list is in Agents. Devbroom is free to download; Pro adds automatic upkeep, Analytics and the session tools.
More guides
- How to clear Xcode DerivedData on a MacWhat DerivedData holds, how to see which project uses the most, Clean Build Folder versus deleting, and the simulators that take even more.
- How to find and clean node_modules and build folders on a MacFind the biggest node_modules and build folders, check that each one can be rebuilt, and move them to the Trash instead of rm -rf.
- How to clean the pnpm store on a MacFind the store, spot the folder an old pnpm version left behind, prune what no project uses, and know why freed space can be less than du says.
- How to free up disk space taken by Claude Code on a MacTranscripts, checkpoints, scratch folders and worktrees: what Claude Code keeps, what you can delete and what you should leave alone.