Apfel
The free AI already on your Mac. CLI tool, OpenAI-compatible server, and interactive chat — all on-device via Apple Intelligence. No API keys, no cloud, no downloads.
Last verified:
What is Apfel?
Apfel is an open-source CLI tool that unlocks Apple's built-in on-device language model on Apple Silicon Macs running macOS 26 (Tahoe). It gives you command-line access to Apple Intelligence with zero configuration—no model downloads, no API keys, no subscriptions. The tool wraps Apple's FoundationModels framework and exposes the ~3B parameter LLM three ways: as a UNIX command-line tool with stdin/stdout, as an OpenAI-compatible HTTP server at localhost:11434, and as an interactive chat with multi-turn conversations.
Key features include native Model Context Protocol (MCP) support for tool calling, pipe-friendly composable design that works with jq/xargs/shell scripts, file attachments, JSON output, proper exit codes, five context trimming strategies, real token counting, and streaming support. The OpenAI-compatible server is a drop-in replacement requiring only a base URL change, working with any OpenAI SDK, LangChain, CrewAI, Claude Code, or frameworks that talk to OpenAI.
Apfel is for developers, DevOps engineers, and Mac users who want private, offline AI for shell scripting, text transformation, automation, quick lookups (math, regex, commands), clipboard operations, and agent runtimes. It's ideal for anyone building local AI workflows, agentic systems, or who values privacy with zero data leaving their machine. The tool has 5.5k+ GitHub stars and is used by engineers from Apple, Google, VMware, NVIDIA, and PlayStation.
Apfel pricing
Pricing model: Freemium
Free forever under MIT license. Zero token costs, no API keys, no subscriptions, no per-token billing. The model is already included with macOS Tahoe. Install via Homebrew: brew install apfel. No payment or sign-up required.
Apfel pros
- Zero token costs—completely free forever
- 100% on-device inference, no cloud required
- Zero configuration—works immediately after brew install
- No model downloads—model already ships with macOS
- No API keys or subscriptions required
- OpenAI-compatible server works with any OpenAI client
- Native MCP support for tool calling with auto-discovery
- Pipe-friendly CLI with stdin/stdout and JSON output
- Works offline—no internet connection needed
- No telemetry, analytics, or tracking of any kind
- Open source MIT license with 550+ tests
- Streaming support with SSE for server mode
- Five context trimming strategies for 4K window
- File attachments supported in CLI mode
- Interactive chat with automatic context management
- Tool calling with automatic execution and result feeding
- Remote MCP servers supported via HTTPS with Bearer auth
- Built-in calculator MCP server as proof-of-concept
Apfel cons
- Only works on Apple Silicon Macs (M1 or later)
- Requires macOS 26 (Tahoe) or newer
- Apple Intelligence must be enabled in system settings
- Limited 4,096 token context window (~3,000 words)
- Fixed model—not configurable or switchable
- Struggles with math and factual recall
- Not suited for long conversations or large documents
- Server mode off by default—must enable manually
- Single model only—no choice of different LLMs
- Does not work on macOS Sequoia or earlier
Frequently asked questions about Apfel
What macOS version do I need?
You need macOS 26 (Tahoe) or newer. The FoundationModels framework only ships with Tahoe and does not work on Sequoia or earlier. You also need Apple Silicon (M1 or later).
Does apfel access my personal data?
No. The FoundationModels framework has zero access to contacts, emails, calendar, photos, or Semantic Index. Those are restricted to first-party Apple features only.
Does any data leave my Mac?
No. The FoundationModels framework works without internet connectivity and runs entirely on the Neural Engine—prompts and responses never leave the device. Apfel adds no analytics, telemetry, tracking, automated update checks, or crash reporting. The codebase is open source and auditable.
What is the context window?
4,096 tokens combined input plus output, approximately 3,000 words. This is great for shell scripts, summaries, and text transforms but not suited for long conversations or large documents.
What is the model good at and where does it struggle?
Good at: shell scripting, text transformation, classification, short summaries, JSON restructuring, and translation. Limited at: math, factual recall, long conversations, and complex code generation. The model prefers to refuse rather than hallucinate—this is deliberate Apple design.
Is the server mode secure?
Server mode is off by default. When enabled, it supports Bearer token auth via --token flag, origin checking, and debug-only log endpoints. Security details are in server-security.md.
Do I need to enable Apple Intelligence?
Yes. Go to System Settings > Apple Intelligence & Siri and enable it. This downloads the model. You don't need to use Siri itself—just enable Apple Intelligence for the model to be available.
How do I update apfel?
Run apfel --update which detects your install method, checks for a newer version, and prompts before upgrading. Alternatively use brew upgrade apfel for Homebrew installs or git pull && make install for source builds.
What model does apfel use?
Apple's own AFM (Apple Foundation Model), roughly 3 billion parameters with mixed 2-bit and 4-bit quantization. It is not OpenAI. The model ships with macOS Tahoe and runs entirely on-device on the Neural Engine.
How is apfel different from Ollama or LM Studio?
Apfel has zero LLM download, instant start, and no setup because it uses the model already on your Mac. Ollama and LM Studio offer more models and larger context windows but require separate downloads and configuration. They are different tools for different needs.