Stop Your Agent From Hallucinating on PDFs

Plain-text PDF tools make agents invent citations. PDF Reader MCP gives them evidence — page numbers, table cells, crops, and provenance.

Install: npm install -g @sylphx/citra

You asked Claude to summarize a 40-page contract. It cited page 12. Page 12 does not say what it claimed.

That is not a model failure. It is a PDF workflow failure.

What actually went wrong

Most agent stacks treat PDFs like .txt files with extra steps:

1. Extract text.
2. Dump it into context.
3. Ask the model to answer.

That pipeline loses tables, skips scanned pages, scrambles multi-column layout, and never tells the agent that content is missing. The model fills the gaps with fluent guesses.

Three failures you cannot fix with a better prompt

Failure	What the agent sees	What actually happened
Table flattening	A paragraph of numbers	Row/column structure was destroyed
Scanned pages	Empty or sparse text	The page was an image, not text
Hidden text	"Extra" content in context	Invisible PDF text layer the human never saw

No system prompt fixes missing evidence.

What evidence-first reading changes

PDF Reader MCP does not stop at text extraction. One read_pdf call returns an Agent Document Twin:

• Markdown and chunks the agent can read.
• Tables with cell geometry so structure survives.
• Document map and AST so headings, lists, and captions keep their roles.
• Trust report so hidden or risky content is flagged, not silently trusted.
• Page + bounding box provenance so answers can link back to source regions.
• search_pdf → pdf_evidence so the agent can verify before citing.

The agent still uses a language model. The difference is it works from claims with coordinates, not from a lossy text dump.

Try the fix in 30 seconds

claude mcp add pdf-reader -- npx @sylphx/citra

{
  "sources": [{ "path": "/absolute/path/to/your.pdf" }]
}

No flags. No provider setup for digital-text PDFs. Inspect the auto_read block to see what route was chosen, then read markdown, tables, and trust signals in one response.

When you need visual proof

{
  "sources": [{ "path": "/absolute/path/to/your.pdf" }],
  "query": "termination clause",
  "max_matches_per_source": 5
}

Take the match page and bounding box into pdf_evidence (render_page or extract_regions) before the agent cites or summarizes.

Why this project exists

Agents are moving from chat to work: contracts, filings, specs, medical records, research PDFs. The cost of a wrong citation is no longer "sounds a bit off" — it is bad decisions on real documents.

PDF Reader MCP is built for that shift: local-first, MCP-native, benchmark-gated, and designed so agents can prove what they read.

If that matches the stack you are building:

• Get started
• See capability comparison
• ⭐ Star the repo so the next builder finds it before they ship another plain-text PDF dump