🍎 Apple Intelligence

The on-device model as a first-class chat provider — private, free, offline.

On a Mac with Apple Intelligence enabled (macOS 26+), this package registers Apple's on-device foundation model as a language-model provider for the agent framework. There is no endpoint to configure and no API key to save — inference runs on the host through Apple's FoundationModels framework, spoken to over a small bundled bridge.

It is the default when nothing else is configured. Cloud providers sort first in the model catalog but only resolve with a key; the on-device model always resolves on a capable host — so a fresh local mesh chats out of the box, and configuring any cloud provider (or picking a model explicitly) takes precedence.

• Private — prompts never leave the device.
• Free — no metered tokens.
• Offline — works without a network.
• Model id: apple-intelligence; on other operating systems the package is a no-op.

Part of the on-device routing direction (platform issue #2432): Gemini Nano on Android follows through the client-executed-round seam.