A Claude Chat Pro–style interface — fully local on Apple Silicon.
oMLX (vision + speed) × Open Interpreter (unrestricted sandbox) × a rich chat UI.
No cloud. No rate limits. Your files, your models, your machine.
oMLX-Interpreter combines three things:
- oMLX — a local MLX inference server built for Apple Silicon, with continuous batching, SSD KV-cache persistence, and vision support. Runs models fast enough for real agent workloads.
- Open Interpreter — an unrestricted code execution sandbox. The model can read, write, and create files on your real filesystem, run shell commands, and iterate on outputs.
- A TypeScript/React frontend — a chat interface styled after Claude.ai, with rich artifact rendering, drag-and-drop attachments, and paste support.
The result: a local AI assistant that can see images, run code, read your PDFs, and write files to disk — with no API costs and nothing leaving your machine.
The closest thing to this was Open WebUI — but Open WebUI requires a Python 3.12 venv and its React dependencies compound the setup complexity. For a local tool you just want to clone and run, that's meaningful friction.
oMLX already solved the hard part: fast, stable local inference on Apple Silicon with vision support. What was missing was a sandbox layer — something that lets the model actually do things locally the way Claude does in the cloud: write files, run code, read your documents, iterate on output. Open Interpreter fills that role without any additional infrastructure.
The --legacy-peer-deps flag in the npm install is the deliberate tradeoff: it sidesteps the React/Python 3.13 dependency conflicts that would otherwise require a venv, keeping setup a single ./setup.sh away.
| Capability | Detail |
|---|---|
| Vision | Image understanding via oMLX's VLM support — paste or drop a PNG/JPEG and ask questions about it |
| Code execution | Open Interpreter runs Python, shell, and more in a real sandbox — not a simulated one |
| File I/O | Reads and writes to your actual filesystem; create, edit, and save files mid-conversation |
| Attachments | Drag and drop or paste: PDF, JSON, Markdown, PNG, JPEG |
| Streaming | Token-by-token streaming responses from the backend |
| Artifacts | Rich rendering of code blocks, markdown, and structured output in the UI |
| Fully local | oMLX runs on-device; no API keys, no data leaving your Mac |
- Apple Silicon Mac (M1 or later)
- macOS 13 Ventura or later
- oMLX installed and running (see below)
- Python 3.13+
- Node.js 18+
Download the .dmg from oMLX Releases, drag to Applications, and launch it. oMLX serves on http://localhost:8000/v1 by default.
Or via Homebrew:
brew tap jundot/omlx https://github.com/jundot/omlx
brew install omlx
brew services start omlxDownload at least one model via the oMLX admin UI at http://localhost:8000/admin.
git clone https://github.com/fxops-ai/omlx-interpreter.git
cd omlx-interpreter
./setup.shsetup.sh installs Python dependencies (including Open Interpreter) and Node packages. Uses --legacy-peer-deps to resolve React dependency conflicts on Python 3.13.
cd backend
uvicorn main:app --host 127.0.0.1 --port 8002 --reloadcd frontend
npm run devOpen http://localhost:3010 — you're live.
omlx-interpreter/
├── backend/ # Python — FastAPI server, Open Interpreter integration
│ └── main.py # API routes, streaming, oMLX proxy, file handling
├── frontend/ # TypeScript/React — chat UI, artifact renderer, attachment handling
├── setup.sh # One-shot install script
├── .gitignore
└── LICENSE # MIT
Language breakdown: ~60% TypeScript (frontend), ~39% Python (backend).
Browser (localhost:3010)
↕ HTTP / SSE streaming
FastAPI backend (localhost:8002)
↕ Open Interpreter (code execution, filesystem)
↕ oMLX API (localhost:8000/v1)
↕ Local model weights (MLX, Apple Silicon)
The backend acts as an orchestration layer: it routes messages to oMLX for generation, passes code blocks to Open Interpreter for execution, handles file attachment parsing, and streams everything back to the frontend in real time.
By default the backend expects oMLX at http://localhost:8000. To point it at a different host or port, set the environment variable before starting:
OMLX_BASE_URL=http://localhost:8000 uvicorn main:app --host 127.0.0.1 --port 8002 --reloadModel selection is handled through the oMLX admin interface at http://localhost:8000/admin. Can also be selected in the chat interface at the user prompt.
MVP — active development.
Core chat loop, streaming, attachments, vision, and code execution are working. The interface matches the intended Claude Chat Pro aesthetic. Rough edges remain — error handling, session persistence, and model switching are still being refined.
If you hit a bug or want to contribute, open an issue.
- oMLX — the inference engine this is built on
- Open Interpreter — the code execution layer
MIT — see LICENSE.
