Scribbler is an open-source AI platform whose core positioning is to let AI run entirely in the browser. It consists of two parts: browser-side AI tools and an AI computer based on ScribblPad, with the goal of running and building AI anywhere, "from the browser to the AI computer."
In terms of use, Scribbler is aimed at developers and users who want to run AI without back-end infrastructure. It runs AI models directly in the browser through WebGPU, with no need to manage servers and no API fees to pay. Data does not leave the user's device, enabling an instant-use experience with no backend, no installation, no npm, and no Python.
Main capabilities include:
- Image generation: Run Stable Diffusion and other diffusion models on-device through WebNN and ONNX Runtime for text-to-image generation, with no API keys or cloud costs, and interactively experiment with prompts and share generated images and notebooks.
- LLM chat: Use WebLLM to chat locally with open-source large models such as Llama, Phi, and Gemma, support building chat UIs and AI agents, text summarization and extraction, completely private, with zero cloud cost and zero cloud latency.
- Machine learning: Use TensorFlow.js, Brain.js, and ONNX Runtime Web to train and run models in the browser, including neural network training, pretrained model inference, classification, regression, and clustering, and visualize training loss and metrics.
- Data visualization: Use Plotly, D3, and built-in tools to analyze datasets and create interactive charts, support loading CSV, JSON, and API data, perform statistical analysis and transformation, and export visualizations as HTML.
- Text-to-speech: Kokoro TTS runs entirely on the client side.
In terms of usage: users can open app.scribbler.live to start immediately, with no installation required; they can also download or clone the GitHub repository to self-host. No npm, Python, or Docker is required, only a browser that supports WebGPU. Notebooks consist of code cells and document cells: code cells run JavaScript, and document cells render HTML/Markdown. AI model outputs (generated images, chat responses, charts) are rendered inline below the cells. Press ▶ or Cmd/Ctrl-Enter to execute, and you can freelyrearrange, add, or delete cells. Libraries such as TensorFlow.js, Transformers.js, WebLLM, ONNX Runtime, Plotly, and D3 can be dynamically imported from a CDN, with no bundler required, loading on demand and being cached. Notebooks can be saved as .jsnb files, shared via URL, pushed directly to GitHub, and notebooks or outputs only can be exported as HTML. The Gallery provides more than 50 AI and data examples, and there are also examples on GitHub; each notebook can be opened instantly in the app via URL.
Applicable scenarios include: browser-side AI experimentation, local private LLM chat, on-device image generation, in-browser machine learning training and inference, data exploration and visualization, interactive teaching and collaborative sharing, as well as air-gapped or enterprise environments. Scribbler is purely static files and can be hosted on any web server, S3 bucket, or GitHub Pages, with no backend process, no database, and no container. For enterprise use and collaboration, contact the official team.
In terms of limitations, the official materials mention that a browser supporting WebGPU is required; information such as the specific supported browser versions, model list, performance limits, and fees and pricing is not yet provided.
