If you own an Apple Silicon Mac and want local models to stand in for cloud APIs, Rapid-MLX is built for exactly that. It serves models through Apple’s MLX framework behind OpenAI- and Anthropic-compatible endpoints, so tools that expect those APIs, including coding agents, can point at your own machine instead.
| Repository | raullenchai/Rapid-MLX |
| Maintainer | raullenchai |
| Primary language | Python |
| GitHub stars | 3,748 (as of 15 September 2026) |
| Last push | 15 September 2026 |
| License status | Apache-2.0 (repository code) |
What it does
According to its README, Rapid-MLX provides:
- A chat REPL and an HTTP server:
rapid-mlx serveexposes an OpenAI-compatible API atlocalhost:8000/v1, with the Anthropic messages route on the same host. - One-command agent setup:
rapid-mlx launch claude-coderewrites Claude Code’s local settings to use the local server, and similar launchers exist for Cline and Continue. - Optional extras beyond text: vision, audio (TTS, STT, voice cloning), video generation, embeddings and local image generation via an OpenAI-compatible Images API (the example uses
flux2-klein-4b). - A macOS desktop app bundling the same engine, for chat, model management and image generation.
Who it’s for
- Mac developers who want coding agents or apps to run against local models.
- People who prefer a one-click desktop app for local chat on Apple Silicon.
Getting started
The README requires an M-series Mac; Windows and Linux desktop builds are not available. Install the CLI with brew install rapid-mlx or the project’s guided installer script, which detects RAM and suggests a starter model. Then rapid-mlx chat starts a chat session with the default qwen3.5-4b-4bit model, a roughly 3GB first download. The base install is text-only at about 460MB; other modalities are opt-in extras. The README links a SECURITY.md for anyone who wants to review the installer before piping it to a shell.
License status
Checked October 17, 2026: The repository code is licensed under Apache License 2.0. Retain the applicable license and attribution notices when distributing it, including the repository’s NOTICE file where required. This does not change the separate licenses of models downloaded through Rapid-MLX.
Limitations and open questions
- Apple Silicon only.
- Its speed claims don’t agree with each other: the GitHub description says “4.2x faster than Ollama”, while the README says “up to 3× Ollama’s throughput (measured)”. Neither has been independently verified by The Signal.
- The README notes Cursor can’t reach a localhost endpoint, because it routes requests through its own servers. Exposing the server publicly to work around this means it is no longer local, and requires setting an API key.
- A young project (created February 2026) with 52 open issues as of 15 September 2026.
Sources
Related: llamafile, single-file local LLMs for any OS · Self-Hosting.
This profile was drafted with AI assistance from the project’s public repository and reviewed by The Signal’s editors before publication. See our AI Disclosure.
