Pixelbrowse
- 9.1k repo stars
- Updated July 31, 2026
- StarTrail-org/PixelRAG
pixelbrowse is a Claude Code plugin skill that screenshots any URL or document with the pixelshot CLI and has Claude read the tiled images visually instead of parsing raw HTML.
About
pixelbrowse is the Claude Code plugin from the Berkeley PixelRAG project (the "Web Screenshots Beat Text for Retrieval-Augmented Generation" paper) that swaps raw-HTML fetching for visual page reading: the skill calls the pixelshot CLI, which renders a URL, local HTML file, PDF, or image to tiled JPEGs through headless Chromium (CDP by default, Playwright as fallback), and Claude reads the tiles as images. The SKILL.md pins the two flags that make captures legible: --tile-height 1568 so Claude's vision model does not downscale text into unreadability, and --wait-network-idle so JavaScript-heavy single-page apps finish rendering before capture. A Pillow crop-and-zoom recipe lets the agent re-read small tables, chart axes, or fine print at full resolution, and a /screenshot slash command wraps the same flow. There is no MCP server or backend - it is a skill-only plugin that runs pixelshot on your machine, giving software engineers chart, diagram, and layout comprehension that HTML parsing throws away.
- Renders URLs, local HTML files, PDFs, and images to tiled JPEGs via headless Chromium (CDP or Playwright)
- Pins --tile-height 1568 so tiles stay within Claude's vision-model limits and text stays readable
- --wait-network-idle waits out JavaScript-heavy single-page apps so captures are not blank or half-rendered
- Pillow crop-and-zoom recipe for re-reading small tables, labels, and chart axes at full resolution
- Skill-only plugin: no MCP server or backend, just the pixelshot CLI on PATH plus a /screenshot slash command
Pixelbrowse by the numbers
- Data as of Aug 5, 2026 (Skillselion catalog sync)
/plugin marketplace add StarTrail-org/PixelRAG/plugin install pixelbrowse@pixelrag-pluginsAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| repo stars | ★ 9.1k |
|---|---|
| Last updated | July 31, 2026 |
| Repository | StarTrail-org/PixelRAG ↗ |
What it does
Ask Claude to look at a staging URL: pixelshot captures it as 1568px tiles and Claude reads the images, catching layout breakage, blank client-rendered views, and chart content HTML fetching misses.
Who is it for?
Software engineers who need Claude to verify rendered UI, read charts or diagrams off a page, or inspect JavaScript-heavy sites that come back blank or garbled from plain HTML fetching
Skip if: Not a browser-automation or interaction tool: it captures static screenshots, so clicking, form filling, and logged-in flows are out of scope. Nor is it the PixelRAG retrieval pipeline - only the capture skill.
When should I use this skill?
When you ask Claude to look at a page, screenshot a URL, check the UI, or read a page visually - including charts, diagrams, infographics, and layouts that raw HTML fetching cannot convey.
What you get
Claude captures the page as image tiles sized for its vision model and reads layout, charts, and fine print directly, with a documented Pillow crop step for details too small to read on the first pass.
- Tiled JPEG screenshots at /tmp/pixelbrowse/<sanitized-url>.png.tiles/tile_NNNN.jpg
- Cropped PNG close-ups produced with Pillow for small text and chart details
- A visual read-back of the rendered page in the agent thread
By the numbers
- 1568px tile height required for legible reads (Claude vision downscales long edges past 1568px on Sonnet/Haiku)
- Default viewport width 875, with 1280 suggested for desktop layouts
- 3-step workflow in the SKILL.md: render, read tile_0000.jpg, read further tiles
README.md
pixelbrowse — Claude Code plugin
Give Claude eyes: screenshot any URL or document with pixelshot and read it visually.
Install
pip install pixelrag # provides the pixelshot command
claude plugin marketplace add StarTrail-org/PixelRAG
claude plugin install pixelbrowse@pixelrag-plugins
Or, for local development from a clone: claude --plugin-dir /path/to/PixelRAG/plugin.
Use
Ask Claude to look at a page:
claude -p "look at https://example.com and tell me what you see"
Or use the slash command in an interactive session: /screenshot <url>.
The skill lives in skills/pixelbrowse/SKILL.md; the command in commands/screenshot.md.
No MCP server or backend — the skill just calls pixelshot (Playwright/CDP) on your machine.
Related skills
TddFollow test-driven development with a strict red-green-refactor loop when creating reliable features or fixing bugs.544k196k
Test Driven DevelopmentEnforce writing failing tests before any production implementation code.181k264kQaRun conversational QA sessions that turn user-reported bugs into well-written, domain-aware GitHub issues without manual ticket writing.179k196kMigrate To ShoehornAutomatically update TypeScript test files that rely on unsafe `as` type assertions by replacing them with type-safe partial objects from @total-typescript/shoehorn.164k196k
Webapp TestingVerify frontend behavior, debug UI issues, capture screenshots, and inspect logs of a running local web application using Playwright.123k165k
Playwright CliRun browser automation, generate element snapshots, inspect DOM attributes, and execute Playwright tests from the terminal.102k12.3kHow it compares
Unlike browser-automation tools that drive a live session through an MCP server, pixelbrowse is a skill-only plugin: one pixelshot CLI call renders the page to image tiles Claude reads with its ordinary Read tool.
FAQ
Does pixelbrowse need an MCP server or an API key?
No. It is a skill-only plugin with no MCP server or backend; the skill just calls the pixelshot CLI (Playwright/CDP) on your machine.
Why does the skill insist on --tile-height 1568?
Claude's vision model downscales images whose long edge exceeds 1568px on Sonnet/Haiku (2576px on Opus), so the default 8192px tiles would turn text unreadable. 1568px tiles keep text legible when Claude reads them.
What can pixelshot capture besides web pages?
URLs (http/https), local HTML files, PDFs, and images. Multiple URLs render in parallel with --workers, --viewport-width 1280 switches to a desktop layout, and the default backend is CDP with Playwright as an option.