Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/StarTrail-org/PixelRAG/llms.txt

Use this file to discover all available pages before exploring further.

The pixelbrowse skill ships with PixelRAG and gives Claude Code the ability to screenshot pages and read them visually — seeing charts, diagrams, tables, and layout the way a person does. Instead of fetching raw HTML that strips away visual structure, Claude uses pixelshot (Playwright/CDP) to capture a page as screenshot tiles and then reads the images directly. No MCP server and no backend are required: everything runs on your local machine.

​
Installation

1

Install pixelshot so it's on PATH

Use uv tool or pipx so pixelshot is available globally and Claude can find it:
uv tool install pixelrag   # or: pipx install pixelrag
A plain pip install into a project virtual environment may leave pixelshot off PATH, which means Claude won’t be able to find it. Use uv tool or pipx to keep it isolated yet always available to Claude.
2

Install the Claude plugin

Add the PixelRAG marketplace entry and install the pixelbrowse skill:
claude plugin marketplace add StarTrail-org/PixelRAG
claude plugin install pixelbrowse@pixelrag-plugins

​
Usage

Once installed, you can ask Claude to look at any page — no extra configuration needed. Non-interactive (CLI flag): 
claude -p "screenshot https://news.ycombinator.com and summarize the top stories"
claude -p "screenshot https://arxiv.org/abs/2404.12387 and explain the key findings"
Interactive session slash command: 
/screenshot https://example.com
The skill just calls pixelshot (Playwright/CDP) on your machine — no MCP server, no backend.

​
How it works

Instead of fetching raw HTML, Claude runs pixelshot to capture a page as screenshot tiles. Claude then reads the images directly — seeing the rendered layout, charts, tables, and diagrams just as a person would, rather than a stripped-down text representation. The pixelrag.md skill file defines exactly when to trigger the tool (on phrases like “screenshot this URL”, “capture this page”, “search Wikipedia visually”, etc.) and which tools are available:
ToolWhat it does
renderCapture a URL or local file to tiled JPEG screenshots
searchQuery a running PixelRAG search API
build indexCreate a searchable visual index from documents

​
Visual search with pre-built indexes

If you have a PixelRAG search API running locally, Claude can query it directly. Send a search request against the running API: 
curl -s -X POST http://localhost:30001/search \
    -H "Content-Type: application/json" \
    -d '{"queries": [{"text": "YOUR QUERY"}], "n_docs": 5}'
The API returns JSON with scored hits including URL, article ID, and relevance score. The following endpoints are available when running local indexes:
EndpointIndex type
:30001Wikipedia text chunks (15.7M vectors)
:30002Wikipedia pixel screenshots (28M vectors)
:30003Wikipedia LoRA+ViT pixel (28M vectors)

​
Build a local index for Claude

You can build a visual index from any local document collection and serve it so Claude can search it. Create a config file, build the index, then start the search API: 
cat > pixelrag.yaml << 'EOF'
source:
  type: local
  path: ./my_docs
embed:
  model: Qwen/Qwen3-VL-Embedding-2B
  device: cpu
output: ./my_index
EOF
pixelrag index build --config pixelrag.yaml --limit 100
pixelrag serve --port 31337
Once pixelrag serve is running, Claude will automatically use it when you ask questions about your documents.
Combine screenshot capture with visual search for full visual RAG workflows: have Claude screenshot a set of pages with pixelshot, build an index from the tiles with pixelrag index build, then query that index to retrieve the most visually relevant tiles before asking Claude to answer questions from the images.

Build docs developers (and LLMs) love

Get started for freeTalk to us