M-Courtyard

Zero-code local LLM fine-tuning & data prep on Apple Silicon. Privacy-first, powered by MLX.

English | 简体中文

Why M-Courtyard?

M-Courtyard is a desktop assistant designed to demystify LLM fine-tuning. Forget about writing Python scripts, managing CUDA dependencies, or renting expensive cloud GPUs. If you have an Apple Silicon Mac, you can build your own custom AI locally.

• Zero-Code Pipeline: From raw PDF/DOCX files to local datasets, MLX fine-tuning, and exportable local runtimes in 4 easy steps.
• 100% Local & Private: No data leaves your machine. Perfect for fine-tuning on sensitive enterprise data or personal journals.
• Optimized for Apple MLX: Powered by mlx-lm, maximizing the potential of unified memory on M1/M2/M3/M4 chips.
• AI-Powered Data Prep: Automatically turn unstructured documents into high-quality instruction datasets using local models, or fall back to built-in rules when you do not want AI generation.

Latest Update (v0.5.9)

• Reliable mlx-lm Detection: Settings and Dashboard now read the installed mlx-lm package metadata instead of importing its full runtime, preventing a valid installation from being shown as missing.

• Setup Verification: Environment setup now verifies the installed mlx-lm version before reporting success.

• uv Environment Setup Compatibility: The managed Python environment setup now rebuilds incomplete or existing .venv directories correctly with newer uv versions.

• macOS Tahoe + MLX Training Stability: M-Courtyard now automatically sets AGX_RELAX_CDM_CTXSTORE_TIMEOUT=1 for training subprocesses to mitigate the upstream MLX / macOS Tahoe Metal watchdog regression that can crash LoRA runs with kIOGPUCommandBufferCallbackErrorImpactingInteractivity.

• Clearer Recovery Guidance: Smart Alerts now recognize this Metal watchdog signature and explain the fallback path if it still appears on Tahoe.

Features

Automated Data Preparation

• Multi-format Import: Drag & drop .txt, .pdf, .docx.
• Smart Segmentation: Automatically clean and chunk documents.
• AI Dataset Generation: Use local Ollama models to generate Knowledge Q&A, Style Imitation, or Instruction Training datasets.
• Built-in Rules Mode: Generate datasets without any external runtime when you prefer a fully self-contained workflow.

Effortless Fine-tuning (LoRA)

• Unified Model Hub: Auto-detect local HuggingFace / ModelScope / Ollama assets, or pull the latest models online (Qwen, DeepSeek, GLM, Gemma, Llama, GPT-OSS, etc.).
• Live Visuals: Real-time training loss charts, ETA, and resource monitoring.
• Presets: 1-click configurations (Quick / Standard / Thorough) for different needs.

Test & Export

• Built-in Chat: Test your fine-tuned adapter instantly.
• One-Click Ollama Export: Merge, quantize (Q4/Q8/F16), and export straight to Ollama. Play with your model immediately.
• MLX Export for Local Runtimes: Export fused MLX models that can be used with mlx-lm.server and loaded in LM Studio on Apple Silicon.

Local Runtime Support

• mlx-lm is the core engine: training and built-in inference are powered by Apple MLX rather than Ollama.
• Ollama is currently optional but recommended: it is used for Ollama-based AI dataset generation and one-click Ollama export.
• LM Studio is supported as a parallel local runtime: use its local OpenAI-compatible server for AI dataset generation, or load exported MLX models there on Apple Silicon.
• Built-in rules remain available with no extra runtime: if you do not want to install Ollama or LM Studio, you can still generate datasets with the built-in rules path.

Interface Tour

1. Data Preparation

Import documents, auto-clean, and generate training datasets using local LLMs.

 

2. Model Training

Real-time loss curves, ETA, and progress tracking powered by Apple MLX.

 

3. Testing & Export

Instantly chat with your fine-tuned model and export it either to Ollama or as MLX assets for LM Studio / local MLX workflows.

 

Requirements

• OS: macOS 14+ (Sonoma or later)
• Chip: Apple Silicon (M1 / M2 / M3 / M4 series)
• RAM: 16 GB+ recommended (for 7B/8B models); 8 GB works for small models (1.5B/3B)
• Core Runtime: M-Courtyard guides the local uv / Python / mlx-lm setup inside the app
• Optional Local Runtime: Ollama installed and running if you want Ollama-based AI dataset generation or Ollama export
• Optional Local Runtime: LM Studio if you want LM Studio-based AI dataset generation or to load exported MLX models there
• No extra runtime required: the built-in rules path can generate datasets without Ollama or LM Studio

Quick Start

Download the Pre-built App (Recommended)

1. Go to Releases and download the latest .dmg.
2. Open the .dmg and drag M-Courtyard.app to your Applications folder.
3. Open Terminal and run this command to allow the app to run (since it's not code-signed yet):

   sudo xattr -rd com.apple.quarantine /Applications/M-Courtyard.app
   

4. Launch M-Courtyard from Applications!

# 1. Clone the repo
git clone https://github.com/Mcourtyard/m-courtyard.git
cd m-courtyard/app

# 2. Install dependencies
pnpm install

# 3. Development mode
pnpm tauri dev

# OR: Production build
pnpm tauri build

Tech Stack

• Frontend: React 19 + TypeScript + TailwindCSS v4 + Vite + Zustand
• Desktop Framework: Tauri 2.x (Rust)
• AI Core: mlx-lm (Apple MLX), local Python venv managed automatically
• Storage: SQLite + local filesystem

Community & Support

Join our community to share your fine-tuned models, get help, or suggest features!

• Discord — Live chat & support
• GitHub Discussions — Feature ideas and Q&A
• GitHub Issues — Bug reports

If M-Courtyard helps you build your local AI, please consider giving it a star on GitHub!

Support

If M-Courtyard saves you time, consider buying me a coffee — it helps keep the project alive! ☕

Chinese supporters can also use 爱发电 (WeChat Pay / Alipay supported).

License

M-Courtyard is open-source software licensed under the AGPL-3.0 License. For brand name and logo usage, see Brand and Logo Usage Notice. For commercial use or different licensing terms, please contact: tuwenbo0112@gmail.com