RemoveMacAI strips Apple Intelligence models off macOS 27
An open-source command-line tool turns Apple Intelligence off, deletes the models it leaves on disk and stops macOS 27 downloading them again.
What the tool does
RemoveMacAI is a small open-source command-line utility for macOS 27. It switches Apple Intelligence off, removes the on-device models that macOS keeps after the features are disabled, and blocks the system from downloading them again. The project says every change it makes can be reverted.
The tool exists because macOS 27 no longer offers one switch for the whole feature set. Apple Intelligence is now split across separate settings panes, and turning those off leaves the downloaded models sitting on the drive. RemoveMacAI opens System Settings to install a configuration profile, which macOS requires the user to approve, then clears the models.
How you run it
Installation is a one-line shell script or the Homebrew tap `omlahore/tap/removemacai`. The script downloads the latest release, checks its SHA-256 checksum and runs it from a temporary directory rather than installing anything. Running `removemacai status` lists each feature and the size of the models on disk, and `removemacai off --keep <features>` leaves the features you name alone.
Release builds are produced by GitHub Actions from their tags and ship with a build provenance attestation, so a download can be traced back to the repository's own source with the GitHub CLI. Homebrew users get the same binary through the tap.
Our opinion
It is telling that the most practical answer to a baked-in AI feature is a stranger's shell script. Apple spent a generation arguing that its assistants were private by design, then shipped a version of macOS where the only honest way to remove the models is a third-party tool that fights the operating system on your behalf.
RemoveMacAI is not a fix for the underlying problem. Every macOS point release can change how the configuration profiles behave, and a tool that runs with enough privilege to strip system models is a serious thing to pipe from the internet into bash. Read the script, verify the attestation and keep it in a VM you care less about — but the fact that people are reaching for it at all is the real story.