M-Courtyard for MAC
Free, open-source Mac app for fine-tuning AI language models on Apple Silicon without code: prepare data, train with MLX, test and export.
M-Courtyard Overview
M-Courtyard is a desktop app for fine-tuning large language models locally on an Apple Silicon Mac, without writing code. You import your documents, turn them into a training dataset, run a LoRA fine-tune with Apple's MLX through mlx-lm, test the result in a built-in chat and export it. The project says no data leaves your Mac.
M-Courtyard is free and open source under the AGPL-3.0 license. It needs macOS 14 or later and an Apple Silicon chip, so it does not run on Intel Macs. The project recommends 16 GB of memory or more for 7B and 8B models; 8 GB works for small models of 1.5B to 3B. The app guides you through setting up uv, Python and mlx-lm. Ollama and LM Studio are optional. The app is not code-signed, so macOS may report it as damaged until you run the command sudo xattr -rd com.apple.quarantine /Applications/M-Courtyard.app in Terminal.
Key Features
M-Courtyard covers the fine-tuning steps from raw document to usable model.
- Document import: drag in .txt, .pdf and .docx files.
- Smart segmentation: cleans and splits documents automatically.
- AI datasets: generate Knowledge Q&A, Style Imitation or Instruction Training data with local models.
- Built-in rules: make datasets without Ollama or LM Studio.
- Model hub: finds local Hugging Face, ModelScope and Ollama models, or downloads new ones.
- Live training view: loss charts, time remaining and resource monitoring, with Quick, Standard and Thorough presets.
- Built-in chat: test a fine-tuned adapter straight away.
- Export: send a quantized model to Ollama, or fused MLX models to local runtimes such as LM Studio.
What’s New in Version 0.5.9
M-Courtyard 0.5.9 improves setup checks and training stability on macOS Tahoe.
- Settings and Dashboard read the installed mlx-lm package details, so a valid install is no longer shown as missing
- Setup now verifies the installed mlx-lm version before reporting success
- The managed Python environment rebuilds incomplete or existing .venv folders correctly with newer uv versions
- Training sets a workaround for the MLX and macOS Tahoe Metal watchdog problem that can crash LoRA runs
- Smart Alerts recognize that crash and explain the fallback
M-Courtyard FAQ
Is M-Courtyard free?
Yes. M-Courtyard is free and open source under the AGPL-3.0 license.
What does M-Courtyard do?
It fine-tunes large language models on your Mac without code. You prepare a dataset from your documents, run LoRA training with MLX, test the model in a chat and export it.
Which macOS version does M-Courtyard need?
macOS 14 Sonoma or later.
Does M-Courtyard run on Intel Macs?
No. The project requires Apple Silicon, and the download is built for it only.
How much memory do I need?
The project recommends 16 GB or more for 7B and 8B models. 8 GB works for small models of 1.5B to 3B.
Why does macOS say M-Courtyard is damaged?
The app is not code-signed. After moving it to Applications, the project says to run sudo xattr -rd com.apple.quarantine /Applications/M-Courtyard.app in Terminal.
Do I need Ollama or LM Studio?
No. Both are optional. They add AI dataset generation and export targets, while the built-in rules can create datasets without them.
Which files can I use for training data?
You can import .txt, .pdf and .docx files.
Does my data leave my Mac?
The project says no data leaves your machine. The app can download models from online sources if you choose to.
Where does the M-Courtyard download on m-courtyard.FULLMAC.app come from?
The download button leads to the official M-Courtyard releases page at https://github.com/Mcourtyard/m-courtyard/releases/latest. m-courtyard.FULLMAC.app does not change or repackage the app.
Similar Apps
LlamaportLlamaport is a small Mac app for running and downloading local GGUF language models. It wraps llama-server from the llama.cpp project: it lists the models in your models folder, launches one with the full command shown before you press start, shows what it costs while it runs, and fetches new models from Hugging Face with downloads that can resume after a kill or a reboot. It reads GGUF headers directly, so the list shows the real quantisation, context length and mixture-of-experts structure rather than guesses from the file name.AI
UnfoldlyUnfoldly is a local AI search app for Mac that finds files by what you remember about them, not by their names. You choose the folders and files it should index, and it builds a searchable memory of them on your Mac: text from documents, what is shown in images, moments in videos and the content of audio recordings. Then you describe what you are looking for, such as a photo of your dog or the PDF that mentioned a pricing plan, and ask follow-up questions about the files it finds: summarize a document, ask about a video or compare notes.AI
Claude Usage MonitorClaude Usage Monitor is a free, open-source menu bar app that shows how much of your Claude.ai usage you have left. Its menu bar icon turns from green to orange to red as you approach your limits, and the text next to it shows either your current session percentage or, while you are working, an estimate of the time left at your present pace together with your weekly usage. Click it to open a small dashboard with progress bars for the current session, your weekly limits across all models and, on Max plans, Sonnet-only usage, each with a countdown to its next reset.AI
EnviousWisprEnviousWispr is a dictation app for the Mac menu bar. Press a keyboard shortcut in any app, speak, and your words are transcribed on your Mac and pasted where you were typing. You can choose between two speech models: Parakeet Ultra, the fast default for 25 European languages, and Whisper, which covers more than 99 languages and detects the language automatically. It works offline and needs no account or subscription.AI
Qwen ScribeQwen Scribe is a menu bar app with a local web interface that transcribes audio and video files and adds system-wide dictation on Apple Silicon Macs. It runs the Qwen3-ASR speech model on the Mac's own GPU through MLX. There is no account, no API key and no Qwen Scribe cloud service, so audio and transcripts stay on the Mac.AI
Agent Teams AIAgent Teams AI is a desktop app for coordinating teams of AI agents, for coding and broader project work. You create a team with one prompt and give its members roles. The agents then work in parallel, message each other, manage their own tasks and review each other's work, while tasks move across a live kanban board. You can send a message to any agent, comment on a task, and approve or reject the code each task changed.AI
FULLMAC.app does not host this app. Download buttons lead to the developer’s own pages.