apfel exposes Apple Intelligence locally

Arthur Ficial’s apfel project packages Apple’s on-device language model into a command-line tool and a local HTTP server that mimics the OpenAI API, according to the project’s GitHub repository and website. The software is positioned as a way for Mac users to send prompts to Apple Intelligence locally rather than through a cloud API. Apple’s own developer documentation shows the underlying Foundation Models framework is now available to app makers on macOS 26 and other current operating systems. Apple says the Foundation Models framework gives developers direct access to the on-device large language model at the core of Apple Intelligence. The company also says the framework supports offline use, guided generation and tool calling. apfel sits on top of that layer and exposes it in forms developers already use: terminal commands, a chat interface and an OpenAI-style endpoint. ### How does apfel actually connect to Apple Intelligence? GitHub documentation for apfel says the tool uses Apple’s Foundation Models framework, which Apple documents as the API for its on-device model. Apple says that framework powers text generation and can call developer-defined tools to complete tasks. The framework is listed for macOS 26.0 and later, alongside iOS 26.0, iPadOS 26.0 and visionOS 26.0. The apfel repository describes the software as “Apple Intelligence from the command line” and says it wraps the on-device model in both a CLI and an HTTP server. The project materials say inference runs on-device and makes no network calls. ### What does “OpenAI-compatible server” mean in practice? The repository says apfel can run as a local server, letting other applications send requests to a Mac over an API format designed to resemble OpenAI’s. That matters because many developer tools, scripts and agent frameworks already expect that interface. Project materials also say apfel supports standard command-line use, including piping text from other Unix tools. That gives developers a way to slot Apple’s local model into existing shell workflows without rewriting everything around Apple’s native Swift framework. ### What hardware and software does it require? Apple’s documentation says Foundation Models requires macOS 26.0 or later on supported devices, and says users must have Apple Intelligence turned on. The apfel materials say the tool is for Apple Silicon Macs and does not require separate model downloads. The project website says installation is available through Homebrew with a single command. GitHub pages tied to the project also reference packaged releases and a Homebrew tap, indicating the maintainer is distributing the tool in a form aimed at developers rather than only as source code. ### Is this the first project to expose Apple’s model this way? GitHub search results show other developers have already built OpenAI-compatible wrappers around Apple’s on-device model. Examples include repositories that describe themselves as local API servers for Apple Intelligence or Foundation Models. apfel’s pitch is narrower and more user-facing. The project materials emphasize no API keys, no subscriptions and no per-token billing, and present the Mac itself as the model host. Franz Enzenhofer’s apfel site describes the software as a way to use “the free AI already on your Mac,” while the underlying GitHub repository credits Arthur Ficial’s codebase. ### What is Apple itself offering developers? Apple’s developer site says developers can “tap into the model with as few as three lines of code” through the Foundation Models framework. The company says the framework can be used for summarization, text extraction, structured output and tool calling, and that experiences built on it can work without internet connectivity. Apple also says Shortcuts now includes a “Use Model” action that can access Apple Intelligence models on-device or through Private Cloud Compute, depending on the task. That means apfel is arriving alongside Apple’s own broader push to make its model accessible inside apps and workflows, though Apple’s official path remains its native framework. Apple’s documentation and the apfel repository both point to the next practical step: developers can install the tool on Apple Silicon Macs running macOS 26 or later and test it against local CLI or OpenAI-style calls. GitHub activity on the apfel project shows recent commits, releases and Homebrew-related packaging work as the software continues to be updated.