Apfel

Apfel exposes the language model built into Apple Silicon Macs (Apple FoundationModels, the model behind Apple Intelligence) as a UNIX command-line tool and as a local OpenAI-compatible server. Inference runs entirely on the device, with no API keys and no network calls. An interactive chat mode serves for testing prompts and MCP servers.

As a command-line tool it takes prompts as arguments or piped input, attaches files (text, PDFs, and images, with on-device OCR), streams output, and returns plain text, code only, JSON, or schema-valid JSON through guided generation. As a server it acts as a drop-in backend for OpenAI SDKs, supporting chat completions, the Responses API, streaming, tool calling, JSON schema output, and image input on macOS 27.

The on-device model has about 3 billion parameters and is not configurable. The context window covers input and output combined: 4096 tokens on macOS 26 and on M1/M2 Macs, and 8192 tokens on macOS 27 with an M3 or newer chip and 12 GB or more of memory. Embeddings and legacy text completions are not available, Apple's guardrails may block benign prompts (a permissive mode reduces false positives), and the model has no knowledge of the current date.

A family of companion projects builds on it, including a macOS chat client, clipboard and overlay utilities, MCP servers sized for the small context window, a TOML configuration wrapper, and a Swift package for tools that talk to the server.