tauri-apple-intelligence

Native Tauri commands for Apple Intelligence (Foundation Models) with streaming, tool calling, Private Cloud Compute, reasoning, multimodal image input, and live token/context usage.

Companion JS transport: https://github.com/entro314-labs/apple-intelligence-sdk

Install

[dependencies]
tauri-apple-intelligence = "0.6"

Capabilities

Capability	Command(s)	Notes
Availability (on-device)	apple_ai_check_availability	Device eligible + Apple Intelligence enabled + model ready
Availability (Private Cloud Compute)	apple_ai_pcc_check_availability	macOS 27+; larger, reasoning-capable, still private (no API key/bill)
Generate / stream	apple_ai_generate, apple_ai_stream, apple_ai_cancel_stream	Basic, tools, and structured modes; model: "on-device" | "private-cloud", reasoningLevel, sampling (temperature incl. 0, topP, topK, seed), toolChoice (macOS 27), and per-message images
Context window	apple_ai_context_info	Real contextSize per model (4k on-device, ~32k PCC) — stop hardcoding
Token counting	apple_ai_token_count	Real tokenizer counts (tokenCount(for:), macOS 26.4+) for budgeting prompts against contextSize
Supported languages	apple_ai_supported_languages	Live BCP-47 tags from SystemLanguageModel.supportedLanguages
Prewarm	apple_ai_prewarm	Lower first-token latency; optional promptPrefix eagerly processes a known prompt prefix

Generation results and streams carry a usage object (inputTokens, cachedInputTokens, outputTokens, reasoningTokens) on macOS 27+.

Typed errors

Generation failures carry a stable machine-readable code mirroring the FoundationModels error cases — context-window-exceeded, guardrail-violation, refusal, rate-limited, concurrent-requests, and more. Non-streaming commands reject with { type: "generation", code, message, contextSize?, tokenCount? }; streams emit an error event with the same fields. context-window-exceeded includes the model's contextSize and the offending tokenCount (macOS 27+), so hosts can condense the conversation and retry — Apple's documented recovery strategy for the context window.

Usage

Then register the commands in your tauri::Builder:

tauri::Builder::default()
  .invoke_handler(tauri::generate_handler![
    tauri_apple_intelligence::apple_ai_check_availability,
    tauri_apple_intelligence::apple_ai_pcc_check_availability,
    tauri_apple_intelligence::apple_ai_generate,
    tauri_apple_intelligence::apple_ai_stream,
    tauri_apple_intelligence::apple_ai_cancel_stream,
    tauri_apple_intelligence::apple_ai_context_info,
    tauri_apple_intelligence::apple_ai_token_count,
    tauri_apple_intelligence::apple_ai_supported_languages,
    tauri_apple_intelligence::apple_ai_prewarm,
  ])

For the JavaScript side, use the companion npm package:

import {
  createAppleIntelligenceProvider,
  createTauriAppleIntelligenceTransport,
} from "@entro314labs/apple-intelligence-sdk";

Prebuilt native library

This crate ships a prebuilt libappleai.dylib in the prebuilt/ directory. You must bundle it into your Tauri app resources so the linker and runtime can find it.

Recommended flow in your Tauri app:

1. Copy the prebuilt dylib into your app's src-tauri/resources/ folder during the build.
2. Reference it in tauri.conf.json under bundle.resources.

There are two common ways to automate the copy step:

• App build script: use your app's src-tauri/build.rs (not this crate) to copy prebuilt/libappleai.dylib into src-tauri/resources/.
• Before-build hook: add a small script (Node/Bash) and call it from tauri.conf.json's build.beforeBuildCommand or your package manager's build pipeline.

If you want, the copy script can locate this crate via cargo metadata and copy the dylib from the Cargo registry to your app's resources folder.

Linking the native library

This crate expects a libappleai.dylib (Foundation Models bridge) to be linked into your Tauri app. Make sure the dylib is bundled as a resource and that the linker can resolve it at runtime.

Command prefix

The JS transport expects apple_ai_* command names by default. If you rename the commands or use a different prefix, update the JS transport's commandPrefix option.

Supported platforms

• ✅ macOS 26+ on Apple Silicon (on-device model, streaming, tools, structured output, sampling modes, typed error codes)
• ✅ macOS 26.4+ adds apple_ai_token_count (tokenCount(for:))
• ✅ macOS 27+ adds Private Cloud Compute, reasoning levels, multimodal image input, per-call token usage, native toolChoice enforcement, and context-size details on context-window-exceeded errors — all gated behind @available, so the crate still runs on macOS 26 with those features simply unavailable
• ❌ Other platforms (returns UnsupportedPlatform)

License

MIT