[Proposal] Bringing mlx-coder (On-device coding agent + Vision support) to the MLX ecosystem

[eduardogoncalves]

Hi everyone,

I’ve been building mlx-coder, an on-device coding agent built natively in Swift for Apple Silicon. My goal was to create a tool that leverages MLX for in-process inference, prioritizing privacy and performance by eliminating external servers and the overhead of Node.js or HTTP bridges.

I’m reaching out to see if the community feels this would be a good fit for the ml-explore organization or how I can best contribute to the mlx-swift ecosystem.

Key features:

• Pure Swift & MLX: No Python bridges or heavy dependencies, just native performance.
• Zero-Dependency Architecture: Bypasses the need for local web servers, making it lightweight and truly "local-first."
• Interactive REPL: A chat interface tailored for rapid prototyping and agentic workflows.
• Vision Support (VLM): I’ve recently integrated support for Qwen 3.5 and Gemma 4, enabling multimodal tasks like code generation from visual context.
• Sandboxed & Secure: Leverages native macOS sandboxing to ensure a secure execution environment, providing peace of mind when the agent interacts with local files.

I believe mlx-coder can serve as a practical reference for developers building complex, agentic applications on top of mlx-swift.

I’d love to hear your thoughts on the best path forward. Should I keep this as a standalone community project, or is there interest in bringing this closer to the official ml-explore org?

Best regards,
Itamar Eduardo Gonçalves de Oliveira

[davidkoski]

That looks really cool! I think it is probably too large for mlx-swift-examples -- the goal of this repo is to give people things to start with to make their own projects.

There is this list of community projects:

• MLX Swift Community Projects mlx-swift#152

and there is a discord community that might like to see it:

• MLX Community Discord mlx#733