What is apple-foundation-models-ruby-sdk?

This library lets Ruby programs use Apple's on-device AI model, which Apple ships as part of Apple Intelligence on recent Macs. Because the model runs on the device, no data is sent to any server and no API key is needed. The library is a port of Apple's own Python SDK and connects to the same underlying Swift code through Ruby's FFI interface. To use it, you need a Mac running macOS 26.0 or later with Apple Intelligence enabled, a full Xcode installation (the lighter Command Line Tools are not enough), and Ruby 3.1 or newer. You add the gem to your Gemfile, run bundle install, and a small Swift package compiles during setup. The main interaction is a session: you create one, optionally give it system instructions, and then call respond with a text prompt. The model returns either the full response at once or a streaming sequence of partial results you can print as they arrive. You can attach images to prompts for the model to describe or analyze, and you can set options like temperature and response length. For structured output, you define a Ruby class with typed properties and optional constraints, such as allowed values, numeric ranges, or required array lengths. Passing the class to respond tells the model to fill in those fields and return an instance. Tools work in a similar way: you write a class with a name, a description, and a typed arguments structure, and the model calls it during generation when it decides the tool is relevant. Other features include exporting conversation history as JSON and resuming a session from a saved transcript, cancelling requests with a timeout, and an optional Fiber-based async layer that lets multiple sessions run concurrently. Errors are organized into named subclasses for common failure modes like guardrail violations and rate limits. The project is Apache-2.0 licensed.