Vmlx is an MCP server tool developed by Jjang AI, designed to run large AI models locally on Apple Silicon hardware. Built primarily with Python, it utilizes Apple's MLX framework for optimized performance. Vmlx serves as an open-source local inference server for large language models, vision-language models, and image generation models, operating as an HTTP inference server on macOS. It is available under a free license and categorized as an MCP server and an AI Agents skill, offering direct model execution on M-series chips.

What does Vmlx connect to?

Vmlx functions as an open-source local inference server, connecting to and serving various large AI models including LLMs, VLMs, and image generation models. It is built exclusively for Apple Silicon hardware (M1/M2/M3/M4), running locally on macOS as an HTTP inference server. Its primary function is to enable local execution of advanced AI capabilities, bypassing cloud dependencies. Vmlx is categorized as an MCP server and an AI Agents skill, providing a self-contained environment for AI model operations.

Core features and tools

Vmlx exposes several callable tools and API endpoints. These include Tool Calling with auto-detected parsers for model families like qwen, llama, and mistral, and Reasoning / Thinking Mode with parsers for qwen3 and minimax_m2. API endpoints cover Chat completion (streaming), Anthropic Messages API, Embeddings, Text-to-speech, Speech-to-text, Image generation, and Reranking. CLI commands include "vmlx serve", "vmlx convert", "vmlx info", and "vmlx bench". The MLX Studio desktop app adds Chat, Vision, and Session Management features.

Who is it for?

Vmlx is intended for developers, users of AI coding assistants, and creative professionals seeking to run large AI models locally. Individuals aiming to deploy AI capabilities directly on Apple Silicon hardware will find it beneficial. It caters to those who want to avoid cloud APIs and associated expenses for their AI inference needs. Skip this if you need to deploy AI models on non-Apple Silicon hardware.

How to get started

The server authenticates using an optional API key for specific /v1 endpoints. It primarily supports streamable HTTP as a transport protocol.

Use cases

• Local AI coding assistance: Provides a capable coding assistant with agentic chat mode, built-in tools like file operations, Git integration, and terminal access, functional even offline.
• Multimodal content creation: Chains capabilities to generate image concepts, edit them with natural language, generate audio, and produce video descriptions locally via an OpenAI-compatible API.
• Self-hosted LLM inference: Runs large language models, vision-language models, and image generation models locally on Apple Silicon, eliminating the need for cloud APIs.

Vmlx offers a powerful, local solution for running AI models directly on Apple Silicon, enhancing privacy and control. Its Python foundation and MLX framework use provide optimized performance for macOS users. The tool's ability to host various model types and expose multiple APIs makes it versatile for a range of AI applications.