macOS MLX Control Center v0.4 Released

The simplest, fastest 1-click Web GUI & CLI Control Center to search, download, compare, benchmark, and run local Multimodal Vision & Text LLMs on Apple Silicon M-Series processors.

Designed specifically for macOS, macOS MLX Control Center bridges Apple's high-performance mlx-lm

(text) and mlx-vlm

(vision) frameworks with a dark glassmorphic web dashboard and terminal interface. It turns any Mac powered by Apple Silicon M-Series processors into a local AI power-station with zero configuration hassle.

Search Models

Compare Models

Customize API for Local Coders like OpenCode or Pi Code

Open Terminal and paste this single command:

curl -sSL https://raw.githubusercontent.com/mypbs/mac-mlx-control-center/main/install.sh | bash

• Download or clone this repository to your Mac.

• Double-click in Finder.start.command

• The server launches and automatically opens the dashboard in your default browser ( http://127.0.0.1:9998

)!

Most local LLM runners are either overly complex, hog system menubars, or require manual hand-editing of configuration JSON files for coding tools. macOS MLX Control Center solves this with:

• 👁️ Vision & Multimodal Image Recognition (v0.4): Full native integration withmlx-vlm

! Run Gemma 4, Qwen 2.5/3.8 VL, PaliGemma, Pixtral, LLaVA, SmolVLM, and more with full image analysis, OCR, visual question answering, and chart understanding. - 🖼️ Interactive Vision Test Chat (v0.4): Upload and preview images directly in the GUI test bench to test OpenAI-compatible multimodal payloads. - ⚙️ Dual-Engine Auto-Detection (: Automatically detects model architectures and chooses the optimal inference engine without manual flags.mlx-vlm

+mlx-lm

) (v0.4) - 🔌 LibreChat & Coding Agent Multi-Modal Sync (v0.4): Direct compatibility with LibreChat (host.docker.internal:9999/v1

), Pi Code, and OpenCode for image recognition and coding workflows. - 🧠 macOS Memory & Process Manager (v0.3): Real-time visual breakdown of Apple Silicon Unified Memory (App, Wired, Compressed, Available Cache) with an interactive top process scanner (🔍 Manage RAM

) and 1-click process termination to free up memory for larger models. - 🔤 Model Sorting & Date Added (v0.3): Sort local and downloaded models by Name (A-Z / Z-A), Size (Largest / Smallest), and Date Added (Newest / Oldest) with exact timestamps on all model cards. - 📱 Horizontal List Mode vs. Card Mode (v0.3): Toggle seamlessly between sleek horizontal list rows and spacious square card grid views across all model and search views. - 🟢 Activity Monitor-Aligned RAM Accounting (v0.3): Accurate unified memory metrics matching Activity Monitor,htop

, andmactop

. - 🚀 1-Click Model Launching & Hot Swapping: Instantly switch between MLX models without terminal commands. - 📦 Multi-Quant & Subfolder HF Support: Direct support for repositories with multiple quantizations (e.g.orcarouter/Qwen3.8-27B-Uncensored-MLX/tree/main/4-bit

). Pick exact bit precision (2-bit

,4-bit

,6-bit

,8-bit

) and download only the weights you need instead of 80+ GB! - 🌐 Supercharged Hugging Face Search: Search thousands of model weights directly with instant filter chips (⚡ MLX Models

,👁️ Vision & Multimodal

,🎯 4-Bit Quantized

,💻 Code Models

) and sorting (Most Downloads

,Most Liked

,Recently Updated

). - 🤖 Built-In AI Model Comparison Engine: Select 2 to 4 models to generate an instant technical breakdown comparing quantization precision (2bit-DQ

,OptiQ-4bit

), RAM compatibility, and specialization —works 100% locally out-of-the-box with zero API keys or accounts required! - 🟢 "Can I Run It?" Apple Silicon RAM Calculator: Live color-coded RAM budget badges (🟢 Fits Smoothly

,🟡 High RAM Pressure

,🔴 Swap Danger

) computed against your Mac's actual available memory. - 🤖 Auto-Sync Pi Code & OpenCode Agents: Automatically configuresPi CodeandOpenCodeunder providerMyMac

(http://127.0.0.1:9999/v1

). Whenever you swap models in the GUI, your coding agents automatically route to the newly loaded model! - 📊 Tokens/Second Speed Benchmark: Measure generation speed (tok/s

), execution time, token counts, and GPU memory usage on Apple Silicon. - ⚙️ Custom API Settings: Change default ports (9999

,8888

,8080

), host binding (127.0.0.1

or0.0.0.0

for LAN access across home Wi-Fi), and max output token budgets with built-in memory guidelines.

Running local AI on your Mac has never been easier. You only need:

Any Mac with Apple Silicon M-Series Processors(with Unified Memory).** macOS**(12.0 or newer).** Python 3.10+**(pre-installed on macOS or available via Homebrew/Xcode).

Note:Dependencies likemlx-lm

andhf

CLI are automatically installed on-demand viauvx

. You do not need to pre-install heavy ML environments! Zero API keys or account sign-ups are required.

Prefer working in the terminal? You can also use mlx.sh

directly:

./mlx.sh

./mlx.sh list

: View all downloaded local models and RAM sizes../mlx.sh search <query>

: Search Hugging Face repositories directly from terminal../mlx.sh start

: Interactive launcher with Fast Swap and Concurrent port options../mlx.sh stop

: Emergency kill-switch for running servers.

In the ** ⚙️ API Settings** panel, you can configure your

Max Output Tokens:

Max Output Tokens: Specifies the maximum number ofgenerated response + reasoning tokensthe server produces. It does not limit your input prompt size.RAM Guidelines for Apple Silicon:** 2B – 8B Models***(Gemma 4 4B, Qwen 7B)*: Set to4096 – 8192

(Light KV Cache RAM impact ~0.5 GB).14B – 35B Models*(Qwen 32B, DeepSeek 35B)*: Set to4096

(Moderate RAM impact ~1.5–3 GB).70B+ Models: Set to2048 – 4096

to prevent memory swapping.

The combo of macOS MLX Control Center and ** LibreChat** gives you a completely private, self-hosted ChatGPT-like experience with full multimodal vision support on your Mac.

• 🔒 100% Local & Private: No data leaves your Mac; runs entirely on Apple Silicon GPU & Unified Memory. - 👁️ Full Multimodal Vision: Attach screenshots, PDFs, documents, diagrams, and photos in LibreChat for instant analysis with models like Gemma 4 and Qwen 2.5/3.8 VL. - ⚡ Hot-Swapping: Switch models in MLX Control Center with 1 click; LibreChat dynamically routes to the active model viadefault_model

without restarting containers. - 💰 Zero Cost: No OpenAI/Anthropic API subscriptions needed.

Add the custom MLX endpoint to your librechat.yaml

file:

version: 1.2.0

endpoints:
  custom:
    - name: "MacBook MLX"
      apiKey: "local"
      baseURL: "http://host.docker.internal:9999/v1" # Use http://127.0.0.1:9999/v1 if running LibreChat natively without Docker
      models:
        default:
          - "default_model"
          - "mlx-community/gemma-4-e4b-it-OptiQ-4bit"
          - "lmstudio-community/Qwen3.8-27B-MLX-4bit"
          - "mlx-community/Llama-3.2-3B-Instruct-4bit"
        fetch: true
      titleConvo: true
      modelDisplayLabel: "MacBook MLX"

Note for Docker users:host.docker.internal

allows Docker containers on macOS to seamlessly communicate with the MLX server running directly on macOS.

If running via Docker:

docker compose restart

• Open LibreChat (e.g. http://localhost:3080

orhttp://localhost:3081

). - Select from the model dropdown.MacBook MLX

• Drag & drop images directly into the chat to perform OCR, code transcription, and visual reasoning!

You can use an AI coding assistant (like Google Antigravity, ChatGPT / Codex, Claude Code, or Cursor) to completely automate managing, launching, and integrating this system on your Mac.

Here are ready-to-use prompt templates you can copy and paste into your AI assistant:

"I want to run local AI models on my Apple Silicon Mac using macOS MLX Control Center. Please runcurl -sSL https://raw.githubusercontent.com/mypbs/mac-mlx-control-center/main/install.sh | bash

to set up the repository, start the GUI dashboard on port 9998, and verify that the server is responding."

"Please configure my local LibreChat setup to connect to my MLX Control Center. Updatelibrechat.yaml

with a custom endpoint named 'MacBook MLX' pointing tohttp://host.docker.internal:9999/v1

withfetch: true

and models list containingdefault_model

andmlx-community/gemma-4-e4b-it-OptiQ-4bit

. Then restart the LibreChat docker container and confirm health."

"Please configure my local coding agent (Pi Code / OpenCode) to use my local MLX server as the default provider under the name 'MyMac' athttp://127.0.0.1:9999/v1

with model IDdefault_model

and reasoning tokens enabled."

"Check the status of my MLX server on port 9999 and GUI on port 9998 usingcurl http://127.0.0.1:9998/api/status

. If I'm using a vision model (like Gemma 4), verify that it is running viamlx-vlm

so image uploads work in LibreChat."

When auto-sync is enabled in ** ⚙️ API Settings**, launching any model automatically updates your local coding agent configs:

Pi Code:~/.pi/agent/settings.json

&~/.pi/agent/models.json

OpenCode:~/.config/opencode/opencode.jsonc

Provider Name: MyMac

Model ID: ** default_model** (Auto-Detect Active MLX Endpoint)

Google Antigravity AI: Pair-programmed, engineered, and designed with Antigravity AI.** Apple MLX Team**: Powered by Apple's open-sourcemlx-lm

&mlx-vlm

frameworks.

MIT License — free to use, modify, and share!

source & further reading

github.com — original article Show HN: Bolnee-Chat – Self Hosted Chatbot Integration in Your Business Website Memctl: Versioned memory for your coding agent (Claude.md / AGENTS.md) KV Cache Size Calculator