The repository provides a 100% on-device AI tool for Apple Silicon Macs, exposing a UNIX CLI and an OpenAI-compatible local server with interactive chat and tool-calling support. It emphasizes no API keys or cloud usage and is actively maintained with multiple releases in 2026.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Arthur-Ficial/apfel is a UNIX tool and OpenAI-compatible server that runs entirely on-device on Apple Silicon Macs, using Apple FoundationModels. It offers a CLI interface, an interactive chat REPL, and a local server for OpenAI SDKs. It requires macOS 26+ and Apple Silicon, with no API keys or cloud dependency.
How it works
The project exposes an on-device LLM via a UNIX tool (apfel) and a local HTTP server at http://localhost:11434/v1 for OpenAI-compatible clients. It supports chat completions, responses API, model info, and tool calling, with configurable context window size (4096 tokens on macOS 26; 8192 on macOS 27). It includes a background server, context strategies, and token controls (max_tokens). It can operate with a local MCP (Model Context Protocol) setup and can serve requests with optional permissive guardrails.
Getting started
Requirements: macOS 26 Tahoe+, Apple Silicon (M1+). Install via Homebrew:
brew install apfel
Upgrade:
brew upgrade apfel
Build from source:
git clone https://github.com/Arthur-Ficial/apfel.git && cd apfel && make install
OpenAI API compatibility base URL:
http://localhost:11434/v1
Quick Start examples for UNIX tool and server usage are provided in the README, including how to use --serve, --chat, and file attachments.
Recent releases
Latest versions and changes include:
- v1.9.0 (2026-08-02): refactor server for single-source SSE header and JSON response construction; dependency bumps; docs updates.
- v1.8.4 (2026-07-22): added Fable-5 apfel engineer subagent; dynamic context window notes for macOS 26/27; fix context_window; other fixes.
- v1.8.3 (2026-07-09): doc fixes for CodeCropper, CLI references; various chore and doc updates.
- v1.8.2 (2026-07-09): chat input handling improvements; doc sync.
- v1.8.1 (2026-07-09): CI guard for source changes; fix for JSON argument balancing.
Traction
Stars: 6250 Forks: 240 Open issues: 11
Behind the repo
Not included in the provided facts. OMIT.
Caveats
License: MIT Created: 2026-03-24 Last push: 2026-08-02 Platform limitations include: one model (apple-foundationmodel, ~3B params on-device), not configurable; 4096 tokens on macOS 26, 8192 on macOS 27; on-device speed is described as a few seconds per response; embeddings and vision are not available on-device; certain API endpoints (e.g., POST /v1/completions, embeddings) return 501 or 400 depending on capability; requires macOS 26+ and Apple Silicon; no network/clock access for the model. These details are stated in the README.