apfel

apfel is a free, open-source Swift CLI tool that exposes Apple's built-in on-device language model (shipped with macOS 26 Tahoe via the FoundationModels framework) as a UNIX command-line tool, an OpenAI-compatible HTTP server, and an interactive multi-turn chat. It runs 100% on-device using Apple Silicon's Neural Engine and GPU — no API keys, no subscriptions, no data leaving your machine. Built with Swift 6.3 and licensed under MIT, apfel bridges Apple's raw LanguageModelSession API with practical developer tooling.

• CLI / UNIX Tool — Pipe-friendly with stdin/stdout, JSON output (-o json), file attachments, and proper exit codes; composable with jq, xargs, and shell scripts.
• OpenAI-Compatible Server — Starts a local HTTP server at localhost:11434; supports POST /v1/chat/completions, streaming (SSE), tool/function calling, response_format: json_object, CORS, and standard parameters like temperature and max_tokens.
• Interactive Chat — Multi-turn conversations with automatic context management and five trimming strategies to handle the 4,096-token context window.
• Zero Cost — No per-token billing, no API keys required; the model is already installed on your Mac as part of Apple Intelligence.
• 100% On-Device Privacy — All inference runs locally on the Neural Engine and GPU; nothing is sent to external servers.
• Power Shell Scripts — Bundled demo scripts include cmd (natural language to shell command), oneliner (pipe chains from English), explain, wtd, gitsum, and mac-narrator.
• Drop-in OpenAI Replacement — Any OpenAI SDK or framework can point to http://localhost:11434/v1 and work without code changes.
• Install via Homebrew — Run brew install Arthur-Ficial/tap/apfel then apfel "Hello, Mac!" to get started in seconds; or build from source with git clone and make install.