The AI coding tools on this Mac are using 12 GB in hidden folders
Every AI assistant installs a runtime, a model cache and a session history in a dotfolder nobody looks at. Measured numbers, and what is safe to clear.
This is the newest category of disk usage on a developer's Mac and nothing measures it, because every one of these tools installs into a hidden folder in your home directory rather than into Applications. They are invisible in Finder, invisible in the macOS storage screen, and they are not small.
See yours in one line
du -sh ~/.[a-z]* 2>/dev/null | sort -h | tail -15That lists every hidden folder in your home directory, largest last. On most developer Macs the top of that list is a mixture of package managers, language toolchains and, increasingly, assistants. Read it with one question: can you name the tool that made each folder?
What is inside them, and which parts matter
| What it is | Typical size | Safe to clear |
|---|---|---|
| Downloaded runtimes and binaries | Hundreds of MB to several GB | Yes, redownloaded on next run |
| Model weights, for local models | 2 GB and up per model | Yes, but it is a large download to get back |
| Session and conversation history | Tens to hundreds of MB | Yes, but you lose the history |
| Indexes of your codebase | Depends on the code | Yes, rebuilt on next open |
| Logs | Small, and never rotated | Yes |
The distinction that matters is between a download and a record. A runtime or a model is a download: clearing it costs bandwidth and nothing else. A conversation history is a record, and it is the one thing in these folders that no reinstall brings back.
Local models are the largest single items
If you have run a model locally, that is almost certainly the biggest folder in the list. Model files are measured in gigabytes each, and downloading three to compare them leaves three on disk.
du -sh ~/.ollama/models ~/.cache/huggingface \
~/.cache/lm-studio 2>/dev/null | sort -hMost of these tools have a command to list and remove models, which is safer than deleting files under them. ollama list and ollama rm are the pattern to look for.
The general lesson about dotfolders
AI tools are the current example, but the category is older than they are: command line tools have always installed into hidden folders, and none of them tidy up after themselves. On the measured Mac the same listing also showed a package manager cache at 5.3 GB and a version manager at 4.5 GB.
Two of those have guides of their own: the npm cache and the Node versions nvm keeps. The wider point is that a cleanup that only looks at Applications and Documents is missing the folders that grow fastest on a working machine.
Common questions
Where do AI coding tools store their data on a Mac?
Almost always in a hidden folder in your home directory, named after the tool, rather than in Applications or Library. Running du -sh on the hidden folders in your home directory lists them, and on a Mac measured in September 2026 five such tools accounted for around 9 GB between them.
Is it safe to delete an AI tool's folder?
The downloaded runtimes, model files and indexes are safe, since they are fetched or rebuilt again. Conversation and session history is the part that does not come back, so check for it before clearing the folder rather than after.
How much space do local AI models use?
Several gigabytes each, and they accumulate because trying three models leaves three on disk. Tools that manage models usually provide a list and remove command, which is safer than deleting the files underneath them.
Why do these folders not show in the macOS storage screen?
Because that screen sorts by category rather than by folder, and hidden directories in your home folder fall into System Data with everything else. There is no drill down there, which is why a command line listing or a disk map is the only way to see them.