Two different problems: caches vs. models
AI tools eat disk in two ways that need different treatment:
- Caches, logs and updater leftovers SAFE — regenerable, delete freely.
- Downloaded model weights WARNING — deletable, but re-downloading 40 GB is a real cost. Delete deliberately, never in a blind sweep.
Desktop AI apps (Electron): Claude, ChatGPT
Both keep their data under ~/Library/Application Support/. The regenerable parts:
| App | Safe to clean | Typical size |
|---|---|---|
| Claude | Cache, Code Cache, GPUCache, crash dumps, ShipIt updater leftovers | 0.5–3 GB |
| ChatGPT | Cache, Code Cache, GPUCache, crash dumps | 0.5–2 GB |
What you must not touch: Cookies, Local Storage, Session Storage, IndexedDB, Local State. That's where your login session lives. Deleting them logs you out — a classic mistake of overly enthusiastic cleanup scripts.
CLI agents: Claude Code, Gemini CLI
CLI tools keep state in dotfolders in your home directory:
- Claude Code —
~/.claude/holds session transcripts, shell snapshots and per-project state. Old transcripts and shell snapshots are deletable;.claude.json, credentials and state files are not — they hold your auth and settings. - Gemini CLI —
~/.gemini/with a similar split: temp/cache files safe,settings.jsonand OAuth credentials not.
AI editors: Cursor, Windsurf, Antigravity
All three are VS Code derivatives, so they inherit VS Code's storage layout under ~/Library/Application Support/<Editor>/:
| Folder | What it is | Risk |
|---|---|---|
| Cache, CachedData, GPUCache | Renderer + extension host caches | SAFE |
| CachedExtensionVSIXs | Downloaded extension installers | SAFE |
| Crashpad, logs | Crash dumps and logs | SAFE |
| workspaceStorage | Per-project editor state | WARNING |
| History | Local file history (undo across sessions) | WARNING |
A long-lived Cursor install commonly carries 2–8 GB of the safe stuff. workspaceStorage grows unbounded too (one folder per workspace you've ever opened), but clearing it resets per-project UI state — your call.
Local LLM runtimes: Ollama, LM Studio
This is where the tens of gigabytes hide:
du -sh ~/.ollama/models
du -sh ~/.lmstudio/models # older versions: ~/.cache/lm-studio
Manage models with the tools themselves rather than deleting files:
ollama list
ollama rm llama3:70b
In LM Studio, use the My Models tab. A handful of experiments — a 7B here, a 70B there — easily adds up to 50+ GB of weights you tried once.
DevCleaner knows all of these out of the box
Free menu bar app with dedicated scanners for Claude, Claude Code, ChatGPT, Cursor, Windsurf, Antigravity, Gemini CLI, Ollama and LM Studio — plus Xcode, npm, Homebrew and 10 more classic toolchains. Caches are rated Safe; model weights are rated Warning and never pre-selected; credential and session files are protected by a hard deny-list the cleaner cannot cross, no matter what's selected.
Download DevCleaner — free ↓FAQ
- Where does Ollama store models on a Mac?
~/.ollama/models— blob files, several GB per model. Useollama list/ollama rminstead of deleting files by hand.- Will cleaning AI app caches log me out?
- Not if you stick to cache directories. Sessions live in Cookies, Local Storage and credential files — a safe cleaner treats those as untouchable.
- Can I delete Claude Code's old session transcripts?
- Yes — they're records of past sessions, not state the tool needs. Keep
.claude.jsonand anything named like credentials or state.