RemoveMacAI collected 588 GitHub stars in its first six days, then drew 444 points and 280 comments on Hacker News in a separate snapshot. Those numbers are attention signals, not proof that the utility works on every Mac. They do identify a sharp user complaint: on macOS 27, reclaiming space from Apple Intelligence takes more than turning off a single setting. The new open-source tool assembles the missing master switch from configuration profiles, private asset-service calls and a local download block.
The project arrived on September 29 and had already reached version 0.2.3 by October 3. Its pitch is unusually narrow: turn off Apple Intelligence features, remove their downloaded models and stop macOS from fetching those models again. The repository supports Apple silicon Macs running macOS 27. It does not support macOS 26 or Intel hardware. Every change is designed to be reversible through the same command-line program.
Apple scattered the controls
Apple's own macOS 27 support page tells users to visit several places to disable its AI features. Siri has a control in System Settings. Message summaries live in Messages settings, email summaries and smart replies live in Mail, and notification summaries have another System Settings panel. The page offers no system-wide Apple Intelligence switch and no instruction for deleting the models already downloaded to the machine.
RemoveMacAI turns that checklist into one operation. Its current catalog covers 14 feature groups, including Siri, Writing Tools, Genmoji, Image Playground, ChatGPT integration, app summaries, inline text prediction, Spatial Photos, Photos Clean Up and Xcode predictive code completion. Some have Apple restriction keys. Others require forced preferences, while a few are represented only by the presence of their model assets. That split is visible in the project's feature catalog.
The distinction matters because the tool does more than reproduce visible switches. Apple documents management restrictions for features such as Safari summaries, Mail smart replies and external intelligence integrations in its Mac deployment reference. RemoveMacAI places those restrictions and the required preferences into a system-scoped configuration profile. macOS still asks the user to approve that profile in System Settings, so the command cannot silently install it.
How one command reaches three layers
The first layer is feature state. The generated profile sets Apple's restriction keys to false and forces preferences for controls that lack a corresponding restriction. It also records which features the user chose to keep. The --keep option is useful here because several features share the same foundation-model set. The code keeps a model set whenever any retained feature still needs it.
The second layer is storage. RemoveMacAI asks Apple's asset service to reset five known sets: the foundation models, image and Genmoji models, Spatial Photos assets, Photos Clean Up assets and Xcode's completion models. The implementation dynamically loads UnifiedAssetFramework from macOS's PrivateFrameworks directory and sends reset operations over Apple's Siri asset-service connection. It leaves System Integrity Protection enabled and does not directly delete files under /System.
That private-framework dependency is the project's most important caveat. Apple can change an undocumented class, selector or service response in an operating-system update. RemoveMacAI checks that expected methods and asset types still match before it sends a reset, which reduces the chance of deleting an unknown set. The interface can still stop working until the project catches up. That is an inference from the code's use of a private framework, not an announced Apple compatibility policy.
The third layer stops the disk from filling again. For every removed model set, the profile overrides the MobileAsset download server with https://127.0.0.1:9/removemacai-blocked/. Port 9 is normally closed, so those fetches fail locally. The profile generator marks the profile as removable. Deleting it restores the user's settings, and macOS can download a model again when a feature requests one.
A dry run is the sensible first run
The repository advertises a curl-to-shell installer, but users do not need to begin by piping remote code into Bash. Homebrew installs the prebuilt release binary, and the CLI can inspect the current machine or generate the proposed profile without applying it:
brew install omlahore/tap/removemacai
removemacai status
removemacai off --dry-run
The installer does contain better checks than its one-liner suggests. It downloads a release archive into a temporary directory, compares its SHA-256 digest with the published checksum, runs the binary, then removes the temporary files. Tagged releases are built by GitHub Actions and include a GitHub build-provenance attestation. A checksum confirms that an archive matches the publisher's manifest. It cannot protect a user if the publishing account and its release assets are compromised together. The attestation gives users a separate way to check that an archive came from the tagged repository workflow.
The executable says it makes no network requests and collects no data. The install path plainly uses the network to retrieve the executable, so that privacy claim applies to the running RemoveMacAI binary. Its job still changes system-wide settings and communicates with a privileged macOS asset service. Reading the generated profile from a dry run is more informative than treating the program's small download size as a measure of risk.
Turning off the model can break developer workflows
Apple now treats its on-device model as a platform service. The company's macOS 27 developer overview says the Foundation Models framework gives apps access to the model that powers Apple Intelligence. RemoveMacAI's documentation says apps using that framework and the Use Model action in Shortcuts stop working after the foundation models are removed. Visual Intelligence and natural-language Calendar editing are also affected.
Dictation remains available because its speech models are separate. The tool can also preserve named features, so a developer who needs Writing Tools can run removemacai off --keep writing-tools. The shared foundation model then stays on disk. This is why the project's status command reports features and model sets separately. A switch can be off while an asset remains because another enabled feature depends on it.
The storage figure is deliberately absent from the project's current introduction. Model holdings differ by machine and enabled features, and macOS performs the final file deletion on its own schedule. System Settings may continue counting Apple Intelligence storage after the asset service has released the models. RemoveMacAI 0.2.1 changed its output after the earlier build claimed the space was free too soon, according to the release history.
Four releases reveal the maintenance burden
Version 0.2.3 fixed a more direct failure. On some Macs, including systems updated to macOS 27.0.1, one asset-service status path returned zero bytes for installed model sets. The tool therefore sent no reset request and removed nothing. The fix reads the asset inventory instead and checks whether the corresponding asset folders still contain files. It also changed release builds to carry provenance attestations.
That correction came four days after the first public release. Fast repair is useful evidence that the maintainer is responding, while the bug shows how quickly this technique can drift across a point update. The repository's automated self-test checks catalog consistency, refusal of unknown sets, profile construction and download redirects. Its continuous-integration job runs on a macOS 15 image, whereas the README reports manual testing on macOS 27.0 and says 0.2.3 or later is required for 27.0.1. Users should read that scope literally.
RemoveMacAI has found an audience because Apple's own controls answer the question feature by feature, while users often ask a simpler one: how do I remove the whole local AI stack and keep it off? The next evidence to watch is mundane and measurable: whether macOS 27 updates leave the profile restrictions intact, whether the private asset calls keep working, and whether release notes continue to name tested versions. If Apple restores a master switch with model removal, much of this utility becomes unnecessary. Until then, its rapid rise is a count of demand for a control the operating system does not provide in one place.