Rapid-MLX launch codex: Run Codex or Claude Code on Local MLX Models (Mac)

Want Codex or Claude Code on a local Apple Silicon model? Rapid-MLX’s one-liner is rapid-mlx launch codex (or claude-code). Per their docs it saves a private profile, starts a local server when needed, and --remove drops only Rapid’s profile.

Up front: this box is not Apple Silicon. We did not install or time Rapid-MLX here. Install commands and speed figures come from their README / comparison pages and are labeled as official claims. For picking local models in general, see Best Local LLM.

What Rapid-MLX is

Per the Rapid-MLX GitHub README: an Apache-2.0 local inference server for Apple Silicon, OpenAI- and Anthropic Messages-compatible, focused on reliable tool calling for coding agents. Agent docs: rapidmlx.com/docs/agents/.

Typical commands from their docs:

rapid-mlx launch codex
rapid-mlx launch claude-code
rapid-mlx launch opencode

Also listed: pi, hermes, goose, crush — follow their agents page.

Install (from their README)

curl -fsSL https://rapidmlx.com/install.sh | bash

Then restart the terminal, or:

export PATH="$HOME/.local/bin:$PATH"

Homebrew is also advertised: brew install rapid-mlx. Install your coding agent next, then rapid-mlx launch ….

I saw those commands in the README text; I did not run install.sh on this machine.

launch and –remove

The agents page says rapid-mlx launch saves Rapid’s own private profile, starts a local model when needed, and leaves your existing agent settings untouched. The same command works again tomorrow.

rapid-mlx launch codex --remove

Docs: removes only Rapid’s private Codex profile, not your existing Codex settings. I inferred from that wording: --remove targets Rapid’s private profile, not a full wipe of ~/.codex.

The older rapid-mlx agents --setup path still edits the agent’s normal config when you want that.

Speed: official numbers only

README claim: up to about 4× vs Apple’s MLX (mlx-lm), about 1.5× on a typical task — same model, same weights. Method note on their page: decode on a Mac mini M4 Pro with Qwen3.5-9B 4-bit, greedy, both servers at defaults; details at their mlx-lm comparison. They also publish tok/s tables vs oMLX and LM Studio. None of these are our measurements. We have no Apple Silicon timings here.

vs Ollama (documented facts)

Feature facts from Rapid’s “How it compares” section (sources dated around 2026-09-26 on their README), not our benchmark:

Item	Rapid-MLX (their README)	Ollama (their docs)
License	Apache-2.0	MIT
Mac engine	MLX	GGML/GGUF, plus optional MLX
Local API	OpenAI + Anthropic Messages	Ollama API + OpenAI/Anthropic compat
Agent launch	rapid-mlx launch codex etc.	e.g. ollama launch claude

If you already wire Ollama into Claude Code, debugging is in selected model fixes. Cloud Chinese models: DeepSeek / Kimi / GLM (ZH guide). Hardware sizing: Best Local LLM and VRAM calculator.

FAQ

Is a slow first reply normal?

Their docs say an uncached model must download and load, and the agent’s opening prompt can be long — no fixed first-reply time.

Which model size for 16 GB / 32 GB?

Agents page examples: qwen3.5-4b-4bit around 16 GB; qwen3.5-9b-4bit for 18–23 GB; qwen3.8-27b-4bit at 32 GB+. Follow their hardware guide.

Did we benchmark it?

No. Multipliers and tok/s above are labeled as their official claims.