mlx-code turns Apple’s MLX stack into a minimal local coding-agent loop for Mac
A new Show HN project called mlx-code packages a lightweight Mac coding agent around Apple’s MLX framework, with local inference, prompt caching, tool calling, and a deliberately Unix-first CLI.
mlx-code is an early-stage Mac coding agent built on Apple’s MLX framework and presented as a terminal-first alternative to heavier hosted tools. The repository describes local inference, built-in prompt caching, and robust tool calling, while the project’s Show HN post confirms it is being pitched as a lightweight “backyard shed” coding setup rather than a polished platform. As of this check, the public GitHub repo was updated on May 9 and showed 17 stars and 2 forks, so this is fresh momentum rather than a mature mainstream release.
Key takeaways
- mlx-code is positioned as a local coding agent for Mac built on Apple’s MLX stack.
- The project splits into a local MLX server (
main.py) and an agent harness (pie.py). - The repo says the local server exposes an OpenAI-compatible completions endpoint.
- The CLI includes
mcfor local use,mefor remote-provider use, andmdfor structured log viewing. - The README highlights support for Claude-, Gemini-, Codex-, and DeepSeek-formatted requests plus disk-backed prompt caching.
What it includes
| Component | What the project says it does | Why that matters |
|---|---|---|
main.py | Runs a clean MLX server for Apple Silicon | Keeps the local model path lightweight |
pie.py | Provides the agent harness and tool execution loop | Makes the project usable as more than a plain inference server |
mc CLI | Starts the local agent workflow | Fast path for local experiments |
me CLI | Connects the harness to remote APIs | Lets users keep the same loop with hosted models |
| Prompt caching | Saves KV cache to disk | Useful if you revisit the same code context repeatedly |
Why it matters
The practical appeal here is not “yet another coding agent.” It is the combination of local-first Mac inference, a standard terminal workflow, and a deliberately small abstraction layer. If you dislike browser-heavy agent products or want a setup you can inspect end-to-end, mlx-code is an interesting pattern: keep the loop in the terminal, keep the interfaces composable, and decide yourself when to use local or remote models.
For LinkLoot readers, that makes this less about hype and more about architecture choice. A local MLX server plus a simple harness can be enough for exploratory coding, note-to-code loops, or controlled tool-use experiments without committing your whole workflow to a hosted UI.
What to verify before you act
First, verify hardware fit. Because mlx-code is built around Apple’s MLX framework, it is a Mac-specific story, and your available RAM plus model choice will shape whether the local experience feels fast or frustrating. Second, confirm how much of your workflow truly benefits from local inference versus the remote me mode, because the project explicitly supports both.
You should also treat the project as early-stage: review the repository activity, license, and issue history before using it for anything business-critical. A fresh Show HN launch can be useful signal, but it is not the same as a long maintenance track record.
Practical LinkLoot angle
If you are testing agent workflows on a Mac, one useful experiment is to compare three loops side by side: fully hosted coding agents, a mixed remote-provider CLI, and a local MLX loop like this one. The comparison usually reveals where latency, privacy, cost, and inspectability matter most in your own work.
For a broader shortlist of practical AI agent tooling patterns, LinkLoot’s guide here fits well: /guides/ai-agent-tools
No. The repo documents a local mc mode and a remote me mode for hosted providers.