Clean AI tools data

AI coding tools write a lot to your home folder, and most of them never clean up after themselves:

• Transcripts of every session, often with screenshots and tool outputs, kept for weeks or forever.
• Per-project data keyed by folder path. Every worktree an agent creates gets its own entry, and that entry stays after the worktree is deleted.
• Old versions. Auto-updaters download a new build (200 MB to 600 MB) and keep the previous ones.
• Caches, logs and databases that only grow: Electron caches, debug logs, a Codex logs database that is never pruned.
• Local models (Ollama, LM Studio, Hugging Face) that you pulled once to try.

lu-cleaner finds these with a dedicated scanner (the ai provider) plus fixed locations from the catalog. It separates what is safe to remove from what is your data, and it never proposes credentials, settings, memories or chat databases.

Note

This page covers the data the tools keep in your home folder. The git worktrees that agents create are covered in Clean up AI agent worktrees.

See what your AI tools use

lu-cleaner scan -c ai          # AI tools data
lu-cleaner scan -c ai,ide      # plus Cursor's editor side (old builds, workspace state, extensions)
lu-cleaner catalog -c ai       # the fixed paths lu-cleaner checks, with their risk

Cursor is split across two categories: its agent data is in ai, and its editor caches (the parts it shares with VS Code) are in ide.

lu-cleaner scan -c ai

AI tools  21.4 GB · 36 items · ★ 6.90 GB recommended
  NAME                                                                    SIZE  AGE  RISK      RECO  PATH
  Ollama model · llama3.1:8b                                           4.92 GB  4mo  caution         ~/.ollama/models
  Codex sessions > 30d · 2026-07 (184 rollouts)                        3.40 GB  2mo  caution         ~/.codex/sessions/2026/07/…
  Claude Code sessions > 30d · ~/dev/my-app (96 sessions)              2.10 GB   5w  caution         ~/.claude/projects/-Users-me-dev-my-app/…
  Codex app runtime (node / python / native deps)                      1.60 GB   3w  moderate  ★     ~/.cache/codex-runtimes/…
  Claude Code data of deleted folder · ~/.codex/worktrees/a1b2/my-app  1.20 GB   6w  moderate  ★     ~/.claude/projects/-Users-me--codex-worktrees-a1b2-my-app/…
  Codex logs database (logs_2.sqlite)                                  1.10 GB  now  safe      ★     ~/.codex/logs_2.sqlite
  Cursor · bundled cursor-agent · 2026.08.31-4057e58 (old version)      610 MB   6w  safe      ★     ~/Library/Application Support/Cursor/User/globalStorage/…
  …

Sizes and names above are illustrative. The ★ column is smart select: items lu-cleaner recommends (they are the ones the a key selects).

Clean it

The interactive picker is the best way to review this category, because much of it is a judgement call:

lu-cleaner clean -c ai

Nothing is selected when the picker opens. Press a to select the recommended items, or pick with space. Items with the caution risk (transcripts, models, generated images) are never recommended, and cleaning them asks you to type yes.

For scripts, start with a dry run:

lu-cleaner clean --yes --smart -c ai --dry-run   # the plan, nothing deleted
lu-cleaner clean --yes --smart -c ai             # clean the recommended items

clean --yes leaves out caution items, and items with a warning (an app running, for example), unless you pass --risk caution. The plan lists the warned items under Held back. To remove old Claude Code transcripts on purpose, name their kind:

lu-cleaner clean --yes -k claude-code-old-sessions --risk caution --dry-run

Quit the apps first

Many items are guarded by the app that owns them: Cursor, Codex, the ChatGPT app, Claude desktop, Conductor, Ollama, LM Studio. While the app runs, the item carries a warning ("Cursor is running — quit it before cleaning"), smart select skips it, and the clean skips it. lu-cleaner doctor lists the running apps that block a cleanup.

A good order: worktrees first

Claude Code and Cursor keep per-folder data. lu-cleaner can only call that data an orphan once the folder is gone. Remove finished agent worktrees first, then scan again. Their leftovers then appear as "data of deleted folder" items:

lu-cleaner worktrees        # remove the finished worktrees
lu-cleaner clean -c ai      # their Claude Code / Cursor data now shows up as orphaned

A folder that disappeared recently may come back (a renamed folder, a restored Conductor workspace, a re-created worktree), and transcripts and chats are never regenerated. So this data is caution and never recommended during the first 30 days after its last write. After that, it becomes moderate and is recommended, unless its folder was only guessed from the directory name or is a Conductor workspace.

Claude Code

What	Where	Risk	Smart select
Sessions older than 30 days, one item per project claude-code-old-sessions	~/.claude/projects/<project>/	caution	No
Data of a deleted folder (orphaned project) claude-code-orphan-project	~/.claude/projects/<project>/	moderate once untouched for 30 days; caution before, or when the folder was guessed or is a Conductor workspace	Only when moderate
Old CLI versions claude-code-old-version	~/.local/share/claude/versions/	moderate	Yes
Debug logs (older than 1 day) ai-claude-code-debug-logs	~/.claude/debug/	safe	Yes
Shell snapshots (older than 7 days) ai-claude-code-shell-snapshots	~/.claude/shell-snapshots/	safe	Yes
Telemetry, feature-flag cache, changelog ai-claude-code-telemetry	~/.claude/telemetry/, ~/.claude/statsig/	safe	Yes
Per-session temp dirs and todo lists (older than 7 days) ai-claude-code-session-temp, ai-claude-code-todos	~/.claude/session-env/, ~/.claude/tasks/, ~/.claude/todos/	safe	Yes
MCP server logs ai-claude-code-mcp-logs	~/Library/Caches/claude-cli-nodejs/	safe	Yes
Paste cache (older than 30 days) ai-claude-code-paste-cache	~/.claude/paste-cache/	moderate	When stale
~/.claude.json backups (all but the newest, older than 7 days) claude-code-config-backups	~/.claude/backups/, ~/	moderate	When stale
File checkpoints behind /rewind (older than 30 days) ai-claude-code-file-history	~/.claude/file-history/	caution	No
Plan files (older than 60 days) ai-claude-code-plans	~/.claude/plans/	caution	No

"When stale" means the item is recommended once it has been unused for longer than stale_after, which is 14 days by default. See Configuration.

A shell snapshot is sourced by every command of the session that created it, so a session left open for days still needs its snapshot: only snapshots older than a week are proposed.

Old sessions. Claude Code stores one transcript per session (<id>.jsonl), with a folder next to it for subagent logs and tool outputs. lu-cleaner groups the sessions of each project that have not been written for 30 days. It skips any session listed as alive in ~/.claude/sessions/. Deleting old sessions removes them from claude --resume and the history. Claude Code also prunes transcripts itself after cleanupPeriodDays (30 days by default). If you raised that setting, lu-cleaner lets you choose which projects to trim.

Old CLI versions. The native installer keeps every downloaded build (about 200 MB each). lu-cleaner keeps the version ~/.local/bin/claude points to, the newest one (the updater may have downloaded it already), and any version a running process uses. To keep more versions, set keep_latest.

How orphaned projects are detected

Claude Code names each project folder after the working directory, with every character that is not a letter or a digit replaced by -. For example, ~/.codex/worktrees/a1b2/my-app becomes -Users-me--codex-worktrees-a1b2-my-app. You cannot decode that name reliably, so lu-cleaner does not try to parse it first:

1. It reads the real cwd of each session from its transcript, or from sessions-index.json. Several folders can share one project directory (my-app and my_app encode the same way), so each session is judged on its own folder.
2. If no session names its folder, it looks for a folder on disk whose encoded name matches. Such an item says the folder was guessed for the whole directory and is never recommended.
3. A folder that cannot be checked (unmounted volume, unreadable parent, macOS privacy protection, a logged-out cloud storage provider) counts as "unknown", never as "deleted".

A session becomes an orphan only when its folder verifiably no longer exists, no live session runs there, and nothing was written in the last hour. It is recommended only when its folder was read from the sessions, is not a Conductor workspace (Conductor restores archived workspaces), and nothing was written for 30 days. Sessions of other folders sharing the directory are kept. The auto-memory (memory/) is always kept: lu-cleaner deletes the transcripts around it, never the folder. Right before deletion, it checks again that the folder is still missing and that no session has started there.

Claude desktop

What	Where	Risk	Smart select
Cowork VM image (rootfs.img, initrd, vmlinuz) ai-claude-desktop-vm-image	~/Library/Application Support/Claude/vm_bundles/claudevm.bundle/	moderate	When stale
Old bundled Claude Code builds claude-desktop-old-claude-code, claude-desktop-old-claude-code-vm	…/Claude/claude-code/, …/Claude/claude-code-vm/	safe	Yes
Electron caches (HTTP, code, GPU, updater) ai-claude-desktop-caches	…/Claude/Cache, …/Claude/Code Cache…	safe	Yes
Cowork VM session disk (sessiondata.img)	…/claudevm.bundle/	caution	Report only
Agent and Code session state, git shadow repos	…/Claude/local-agent-mode-sessions…	caution	Report only

The VM image is a sparse Linux disk, about 10 GB allocated. Claude desktop downloads it again the next time Cowork starts. The session disk, which holds the files of your Cowork sessions, is only reported. For bundled Claude Code builds, lu-cleaner keeps the newest version and the one named in claude-code-vm/.sdk-version. The VM image and the caches wait until the Claude app is closed.

Codex (CLI and app)

What	Where	Risk	Smart select
Sessions older than 30 days, one item per month codex-old-sessions	~/.codex/sessions/YYYY/MM/	caution	No
Archived sessions, one item per month codex-archived-sessions	~/.codex/archived_sessions/	caution	No
Logs database codex-logs-db	~/.codex/logs_*.sqlite (+ -wal, -shm)	safe	Yes, when Codex is closed
Stale copy of the logs database codex-stale-logs-db	~/.codex/sqlite/logs_*.sqlite	safe	Yes
Logs database repair backups (older than 7 days) ai-codex-repair-backups	~/.codex/logs_*.codex-repair-*.bak	safe	Yes
Repair backups of your data databases (memories, goals; older than 30 days) ai-codex-repair-backups-data	~/.codex/*.codex-repair-*.bak	caution	No
App runtime (node, python, native deps) ai-codex-runtimes	~/.cache/codex-runtimes/	moderate	When stale
Plugin caches and sync clones ai-codex-plugin-cache, ai-codex-plugin-staging	~/.codex/plugins/cache, ~/.codex/.tmp/	moderate	When stale
Catalog caches, shell snapshots, app Chromium caches ai-codex-caches, ai-codex-shell-snapshots, ai-codex-app-caches	~/.codex/cache/, ~/.codex/shell_snapshots/, ~/Library/Application Support/Codex/	safe	Yes
CLI log files (older than 14 days) ai-codex-log-files	~/.codex/log/	moderate	When stale
Generated images and visualizations (older than 30 days) ai-codex-generated-images, codex-visualizations	~/.codex/generated_images/, ~/.codex/visualizations/	caution	No
Thread history DB, old state and memories DB copies, PR archives, attachments, ~/Documents/Codex	~/.codex/…	caution	Report only

Sessions. A single rollout file can reach hundreds of MB when it contains screenshots. Deleting one makes that thread impossible to resume. Pinned threads are always kept: lu-cleaner reads their list from the Codex state database with a read-only sqlite3 query.

Repair backups. When Codex repairs a database, it keeps a copy of the file from before the repair. For the logs database that copy is waste. For the memories and goals databases, it is the only way to recover rows a repair dropped, so those copies are only proposed as caution items. Backups of the state database are protected.

Logs database. logs_*.sqlite only holds tracing and feedback logs. Codex never prunes or vacuums it, and creates an empty one at the next start. It is deleted together with its -wal and -shm files, so it must not be open. The item waits until codex, the Codex app and the ChatGPT app are closed (ChatGPT embeds Codex). Right before deletion, lu-cleaner also checks with lsof that no process has the file open.

Cursor

What	Where	Category	Risk	Smart select
CachedData of old app builds cursor-cached-data-old-builds	~/Library/Application Support/Cursor/CachedData/	ide	safe	Yes
Workspace state of deleted folders cursor-workspace-storage-orphans	…/Cursor/User/workspaceStorage/	ide	moderate once untouched for 30 days and without chat data; caution otherwise ("to review")	Only when moderate
Workspace data of uninstalled extensions cursor-workspace-dead-extension-data	…/workspaceStorage/*/<extension>	ide	safe	Yes
Superseded extension versions cursor-extension-old-versions	~/.cursor/extensions/	ide	moderate	Yes
Electron caches, logs, downloaded VSIX ai-cursor-electron-caches, ai-cursor-logs, ai-cursor-vsix-cache	~/Library/Application Support/Cursor/	ide	safe	Yes
Local History (Timeline) older than 180 days ai-cursor-local-history	…/Cursor/User/History/	ide	caution	No
Bundled cursor-agent versions cursor-agent-bundled-old-version	…/globalStorage/anysphere.cursor-agent-worker/agent-cli/	ai	safe	Yes
Standalone cursor-agent and origin CLI versions cursor-agent-old-version, cursor-origin-old-version	~/.local/share/cursor-agent/versions/, ~/.local/share/cursor/origin/	ai	safe	Yes
Agent data of deleted folders cursor-agent-orphan-projects	~/.cursor/projects/	ai	moderate once untouched for 30 days, when Cursor recorded the folder; caution otherwise ("to review")	Only when moderate
Agent MCP descriptor caches cursor-agent-mcp-caches	~/.cursor/projects/*/mcps	ai	safe	Yes
Agent transcripts older than 30 days cursor-agent-old-transcripts	~/.cursor/projects/*/agent-transcripts…	ai	caution	No
State database backups older than 30 days ai-cursor-state-db-backups	…/globalStorage/state.vscdb.backup*	ai	caution: the only restore points of your chats	No

Old builds and versions. Cursor creates a CachedData/<commit> folder for each update, but only uses the commit of the installed Cursor.app, which lu-cleaner reads from the app bundle. The bundled cursor-agent keeps every build it downloads (200 MB to 600 MB each). lu-cleaner keeps only the build its bin/cursor-agent symlink targets and the newest one.

Workspace state of deleted folders. Each folder or worktree you open gets an entry in workspaceStorage. lu-cleaner reads the folder from its workspace.json and proposes the entry once that folder verifiably no longer exists. Remote workspaces are ignored. Entries untouched for 30 days and without chat data form a recommended item. The others (written recently, holding chat history, or of a Conductor workspace that can be restored) go into a separate "to review" item that is never recommended. The agent data in ~/.cursor/projects follows the same split. Just before deletion, lu-cleaner checks again that the folders have not come back.

Extension data. Data of uninstalled extensions (language-server indexes can take gigabytes) is only proposed when lu-cleaner can read the built-in extension list from Cursor.app. Without it, lu-cleaner cannot tell that an extension is gone.

Your chats are never deleted

User/globalStorage/state.vscdb holds every Cursor chat, your settings state and your login. It is protected and only reported, with the never risk. To shrink it, delete old chats inside Cursor. The file only gets smaller after a SQLite VACUUM run while Cursor is closed. lu-cleaner never does that for you.

Other AI tools

Tool	What lu-cleaner proposes	Risk
ChatGPT desktop	Conversation cache (re-synced from the server), HTTP caches	moderate, safe
ChatGPT Atlas	Its whole data folder, only when the app is no longer installed	caution
Conductor	Archived workspace contexts older than 30 days, per repository (notes, plans, often large APKs and recordings)	caution
	Old bundled Claude Code and Codex binaries (keeps the version bin/claude / bin/codex targets)	safe
	conductor.db (sessions and messages) and cache.db (unsent drafts): report only	never, caution
Multica	The codex-home copy of each task completed 7+ days ago. Task workdirs (repository checkouts) are never touched.	moderate
GitHub Copilot CLI, vibe-kanban	Superseded versions	safe
Antigravity	Extensions left after uninstalling the app (moderate). The browser-agent profile left after uninstalling (saved passwords, cookies) and browser recordings older than 90 days (caution, never recommended). Conversations and "brain" are report only.	moderate, caution
opencode, Continue, Windsurf, Raycast, Grok CLI, chrome-devtools MCP	Caches, indexes, logs, downloaded updates and installers. Sessions and chats stay (report only).	safe
VoiceInk	Dictation recordings older than 30 days (transcripts stay in the app)	caution
Gemini CLI, Warp	~/.gemini/tmp (checkpoints) and Warp's database: report only	caution, never

Run lu-cleaner catalog -c ai for the exact paths, or see the catalog.

Local models

Model weights are deliberate multi-gigabyte downloads, so each one is its own item with the caution risk. They are never recommended, and clean --yes ignores them unless you pass --risk caution.

Store	Location	Item
Ollama	~/.ollama/models or $OLLAMA_MODELS	One per model tag. Layers shared with another model are kept.
LM Studio	~/.lmstudio/models, ~/.cache/lm-studio/models	One per model folder
Hugging Face	~/.cache/huggingface/hub, $HF_HUB_CACHE or $HF_HOME/hub	One per model, dataset or space (all revisions)
Whisper	~/.cache/whisper, VoiceInk's WhisperModels	One per model file

lu-cleaner scan -k ollama-model,lmstudio-model,huggingface-model
lu-cleaner clean -c ai          # then press / and type "model" to filter

A model store that is a symlink to another disk is reported, not proposed: deleting it would not free any space on your internal disk. The Ollama identity key (~/.ollama/id_ed25519) is protected.

What is always protected

The safety guard refuses these paths, anything inside them and any folder that contains them, whatever a scanner or catalog entry proposes:

Tool	Protected
Claude Code	~/.claude.json, .credentials.json, settings.json, settings.local.json, CLAUDE.md, skills/, agents/, commands/, hooks/, rules/, installed plugins and marketplaces, history.jsonl, sessions/, every project's memory/
Claude desktop	claude_desktop_config.json, config.json, git-worktrees.json
Codex	auth.json, config.toml, AGENTS.md, memories/, skills/, rules/, automations/, history.jsonl, session_index.jsonl, the global state file and its backup, the state, goals and memories databases
Cursor	~/.cursor/mcp.json, rules/, skills/, User/settings.json, keybindings.json, snippets/, globalStorage/state.vscdb (all chats), storage.json
Others	Ollama identity key, Gemini CLI credentials and settings, ~/.multica, Conductor settings

The tool folders themselves (~/.claude, ~/.codex, ~/.cursor, ~/.conductor, ~/conductor) can never be removed as a whole. To protect more, add paths to protect in your configuration:

~/.config/lu-cleaner/config.toml

protect = ["~/.codex/sessions", "~/.ollama/models"]

These entries are literal paths, not patterns: ~ and $VARS are expanded, and case and Unicode normalization do not matter, as on APFS.

Related

• Clean up AI agent worktrees
• Risk levels and smart select
• What gets scanned
• lu-cleaner clean reference