Local coding agent on a Mac via Rapid-MLX with Codex and Hermes

What they did

They wanted coding agents running fully locally on a Mac, with no cloud API keys. They installed Rapid-MLX on a stock Apple Silicon machine using the curl install script, restarted the terminal so ~/.local/bin was on PATH, and confirmed the rapid-mlx CLI was available. Next they installed their chosen coding agents (Codex and Hermes) as normal, then used `rapid-mlx launch codex` and `rapid-mlx launch hermes` to start each agent against Rapid's private profile, which points the agent at the local OpenAI/Anthropic-compatible endpoint and a local MLX model. They kept their existing agent configs untouched, relying on the private profile so `--remove` could cleanly undo the wiring later. They compared decode speed against mlx-lm and LM Studio on the same 4-bit weights, and checked tool-calling reliability since agent loops depend on it.

How it went

Both agents ran end to end against a local model with no API keys, and decode speed beat mlx-lm by roughly 1.5x on a typical task (up to 4x on some). Tool calling held up well enough for multi-step agent loops. The first reply was slow while the model downloaded and loaded, and larger models needed more unified memory than smaller Macs offer.

Worth knowing

Apple Silicon only, and Python 3.12+ is required. First-run latency includes model download and load. Speed numbers come from single runs on specific hardware, so treat them as directional. You still need to install each coding agent yourself, and the private profile only covers agents Rapid-MLX knows how to launch.

Wiring

Install Rapid-MLX (curl script or brew), ensure ~/.local/bin is on PATH, install the coding agent separately, then `rapid-mlx launch <agent>` starts the local server and points the agent at it via a private profile. The server exposes OpenAI- and Anthropic-compatible endpoints, so agents that speak either protocol work. `--remove` deletes only the private profile.

Try it yourself

Set up Rapid-MLX on this Apple Silicon Mac and run my coding agent against a local model. Steps: 1) install via the official install script, 2) fix PATH, 3) verify `rapid-mlx --version`, 4) pick a 4-bit Qwen model that fits my RAM, 5) launch my agent with `rapid-mlx launch <agent>`, 6) run a small tool-calling smoke test (read a file, edit a file, run a test) and report decode tok/s and any tool-call failures. Do not modify my existing agent config; use the private profile and show me the `--remove` command.