MCP SERVER · OFFICIAL PROJECTPDF answers an agent can cite — page, table cell and bounding box, from a native Rust engine.
LABSCO SUMMARY
Plain-text extraction is what makes agents invent page numbers: once the layout is gone, there is nothing to cite. Citra keeps the geometry, so a figure comes back as a page, a table and a cell you can show someone. The tool surface is deliberately small — advanced work lives inside pdf_evidence's op enum instead of spawning new tool names.
Citra, a local-first PDF reader for agents. It returns structured text, table geometry, OCR output and page-level citations rather than a flat text dump, so a claim can be traced back to a page and a cell. Published as @sylphx/citra, version 5.0.0, binary citra.
read_pdf — the primary read: markdown, tables, document structure, OCR and citations, from a minimal call that only needs the file pathsearch_pdf — literal retrieval that comes back with page and bounding-box locators, for finding the passage before reading deeplypdf_evidence — the follow-up operations behind one op enum: inspect, render, crop, OCR and region extraction@sylphx/citra/sdk, which exposes read, search and evidenceNode.js 22.13.0 or newer for the launcher, and nothing else — no Docker, no API key, no cloud vision service, and PDFs stay on the machine. Run it with npx and the package, or install globally and call citra. One platform native package is selected automatically for your host: macOS arm64 and x64, Linux x64 and arm64 gnu, and Windows x64 are the supported targets. If the matching native binary is missing the server fails closed rather than falling back to a slower engine, which is a deliberate choice — there is no TypeScript PDF runtime in the production package. Optional OCR and visual providers are opt-in and not needed for the core text paths.
One command — npx -y @sylphx/citra
USE CASES
Find the page the figure sits on, read the table structure, and cite the cell by row and column.
Run the OCR path and keep the page-linked evidence, instead of accepting a text soup.
Crop or render the region around the match so the claim and the image travel together.
Read the structure and reading order, then search for the phrase you need with its page locator.
Counted once with only startup setup. Other versions, modes or optional credentials may expose a different list.
pdf_evidenceFocused PDF evidence operations. Pure-Rust supports inspect, bounded page rendering/crops, opt-in bounded command-provider OCR, and region analysis via command, HTTP URL, or ollama/openai-compatible/lmstudio/llamacpp presets.
read_pdfPrimary PDF reader. Pure-Rust default: selectable text, markdown, chunks, tables, search, and evidence-oriented twin fields under capability-first semantic compatibility (ADR-0005).
search_pdfSearches extracted PDF text with page, snippet, geometry, and provenance. Optional OCR uses the bounded command provider.
BEFORE YOU INSTALL
3 tools when we connected. Three tools by design: pdf_evidence handles inspect, render, crop, OCR and regions through an op argument, and read_pdf is the default reading tool.
npx on PATH
CONTEXT COST
Rated by the size of the 3 tool definitions that ride along on every turn, not by how many there are
INSTALLATION
npx -y @sylphx/citraclaude mcp add citra -- npx -y @sylphx/citra{
"mcpServers": {
"citra": {
"command": "npx",
"args": [
"-y",
"@sylphx/citra"
]
}
}
}npm i -g @sylphx/citra
citraMCP servers
labeveryday
hanweg
mehmetozcan-zz
gstiebler