GitHub radar
PixelRAG: Add eyes to your charts and tables with Claude!
A visual retrieval system that turns webpages into images so you can query images instead of text, along with a Claude Code skill that lets your agent see visualizations like a human.
PixelRAG is an open source visual retrieval system developed by Berkeley SkyLab and BAIR. Rather than extracting text from webpages, it turns the webpage (along with other document types such as PDFs and images) into a set of screenshot tiles that can be queried visually. This allows a vision model to answer questions about the webpage’s visual layout, tables, charts, and infographics, which are often lost when the webpage is parsed into text. PixelRAG comes with a pre-built hosted index of 8.28M Wikipedia articles that you can query for free on pixelrag.ai without needing an API key. If you’re using Claude Code, there’s also a pixelbrowse skill available in the repository that provides a /screenshot command. When you run this command, Claude will render the webpage using Playwright, then interpret it visually as an image. This means Claude will be able to read charts and tables like humans do! There is no need for an MCP server or backend, nor any GPUs; the skill simply invokes a local installation of the pixelshot command.
Why a vibe-coder should care
Your Claude agent will never miss another chart, graph, table value, or infographic again. Now, your agent can read webpages like a human, not an HTML parser.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the pixelbrowse skill from https://github.com/StarTrail-org/PixelRAG: run the commands from the 'Give Claude eyes' section of the README (`uv tool install pixelrag`, then `claude plugin marketplace add StarTrail-org/PixelRAG` and `claude plugin install pixelbrowse@pixelrag-plugins`), verify `pixelshot` is on your PATH, and take a screenshot of https://news.ycombinator.com; ask me if you need permissions.
Runs on a regular laptop — Python and Playwright required (both installed together with one command).
Open on GitHub▌ More finds