CitraGive your AI agent eyes for PDFs — with proof.

Zero-config local PDF evidence for agents. `npx -y @sylphx/citra` → structured text, tables, OCR, page-level citations. Plain-text tools make agents guess. Citra returns proof.

Zero-config npx

Star on GitHub

Stop PDF Hallucinations

Product Proof

Benchmark Proof

🚀

Zero-config

`npx -y @sylphx/citra` starts a brand-sole MCP server on stdio. No Docker, no API key, no global install.

👁️

Evidence, not text dumps

Page numbers, bounding boxes, table cells, and provenance so agents can cite instead of invent.

⚡

Measured product proof

Cleaner install than historical TS 3.0.14 on measured linux-x64, and registry-bound dual-mode warm re-reads ≥ ~10× on required fixtures (method-bounded).

📊

Tables agents can trust

Rows, cells, geometry, and confidence for financial reports and structured PDFs.

🔎

Search then verify

Find snippets with page context before deep reading, cropping, or citing.

📄

Agent Document Twin

One linked document map — structure, chunks, layout, trust and accessibility routing when requested.

🖼️

Visual evidence

Render pages and crop regions for tables, figures, charts, and citations.

🔤

Scanned PDF OCR path

Route selected pages through configured OCR providers and keep OCR separate from selectable text.

🛡️

Trust signals

Surface hidden-text, prompt-injection, overlapping, and related risk signals when requested.

⚡

Native local engine

Sole-Rust production on five platforms via a thin Node launcher. Clean install. Fail closed if the native binary is missing.

🔌

Three tools, one surface

read_pdf for smart extraction, search_pdf for cheap evidence, pdf_evidence for focused inspect/render/crop/OCR ops.