M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
P

PDF Reader MCP

Community
An evidence-first PDF intelligence layer for AI agents.
How to install ↓ Compare GitHub source repository ↗
Category
Other #64 of 230
Stars
★ 1.0k Very popular
Transport
stdio (local process)
Runtime
Node.js
Credentials
No credential needed
License
MIT
Last commit
Tools
3
61FMRS · C

The source presents this project as an evidence-first PDF reader focused on document structure, tables, OCR, visual evidence, citations, and trust reporting. It also states that the pure-Rust implementation remains in progress and that the README and manifest contain different version references.

Strongest · Documentation 14/20 Weakest · Reliability 8/20

Reliability
8/20
Security and permissions
14/20
Maintenance
14/20
Documentation
14/20
Setup experience
11/20
Why each score
Reliability 8/20
The README describes read_pdf, search_pdf, pdf_evidence, and fail-closed behavior, but the manifest only declares a stdio package and provides no tool schemas, initialization-handshake evidence, error-handling implementation, or actual test files. The manifest version 4.1.2 conflicts with the README's stated stable version 3.2.2, reducing confidence in reproducibility and behavioral alignment. Under the static calibration, there is no verifiable CI workflow and committed key-path test coverage, so this cannot exceed 12.
Security and permissions 14/20
The supplied material indicates a local-first design, host/private-IP/size/HTTP controls for URL loading, deployment-configured OCR and visual providers, and request-level restrictions against arbitrary executables or provider URLs. These are positive boundary controls. Points are deducted because no implementation code, default permissions, confirmation flow, logging/data-handling behavior, or enforceable network policy is shown; configured command providers and URL loading may still broaden the data boundary.
Maintenance 14/20
The project is not archived and identifies a maintainer organization, MIT license, npm release path, CI, tests, benchmarks, and GitHub Issues. This provides credible maintenance foundations. Points are deducted because the material contains no commit cadence, dependency-update evidence, vulnerability-response process, maintenance commitment, or issue-response history, and the README and manifest versions are inconsistent.
Documentation 14/20
The README includes installation examples for multiple clients, tool purposes, request and response examples, capability descriptions, provider configuration, security notes, limitations, and links to layered documentation. Coverage is therefore strong. Points are deducted for the 3.2.2 README versus 4.1.2 manifest conflict, capabilities marked provider-ready or partial, and limited evidence about authentication, cost, resource limits, platform differences, and reproducible troubleshooting.
Setup experience 11/20
Claude, generic stdio, Claude Desktop, Docker, and library-import examples make the main setup path understandable. Points are deducted because Node.js >=22.13 is required, the default depends on platform-native optional packages and fails closed when they are missing, cross-platform native binaries are not yet complete, and the TypeScript rollback requires an environment variable. These factors increase installation and client-integration fragility.

Static review · not runListed 2026-07-29

Read the FMRS scoring method →

Fit and risk

What it can accessReads local files

Best for

  • Research and analysis agents that need citable page evidence.
  • PDF workflows requiring tables, layout, and visual evidence.
  • Teams seeking local-first document processing with trust reports.

Not for

  • Workflows that only need a simple plain-text dump.
  • Assuming complete pure-Rust Document Twin capabilities; the README states that pure Rust is still a work in progress.
  • Unspecified remote authentication, write, or collaboration features.

Required permissions

  • The server needs access to read the local PDF paths supplied in requests.
  • URL loading, if used, must comply with configured host, private-IP, size, and HTTP policies.
  • OCR and visual-enrichment providers are configured by the deployment environment; requests cannot select arbitrary executables or arbitrary provider URLs.

Risks and side effects

  • PDFs may contain hostile, misleading, or prompt-injection-like content.
  • Hidden text, nearly invisible text, overlapping text, unsafe links, and redaction signals may affect extraction or trust assessment.
  • OCR, table extraction, and visual analysis can produce low-confidence results and should be checked against rendered pages or crops.
  • The README and server manifest identify different product versions, so the intended version should be confirmed before installation.

Setup

Before you start

Runtime：Node.js

MCP_API_KEY optionalsecret Only needed for the opt-in HTTP transport when binding beyond loopback; local stdio usage requires no key.

Add the following configuration to Claude Desktop's claude_desktop_config.json, using npx to launch @sylphx/[email protected]. The server runs over stdio. Other MCP clients can use the same npx command. The README also provides a 3.2.2 stable-release example; the current server manifest lists version 4.1.2.

claude_desktop_config.json 
{"mcpServers":{"pdf-reader":{"command":"npx","args":["@sylphx/[email protected]"]}}}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "pdf-reader": {
      "command": "npx",
      "args": [
        "@sylphx/[email protected]"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add pdf-reader -- npx @sylphx/[email protected]

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After configuring npx -y @sylphx/citra and restarting the client, confirm read_pdf, search_pdf, and pdf_evidence appear in the tool list; call read_pdf on a local PDF and a structured response with page and bbox locators confirms the connection.

Troubleshooting

  1. Confirm that the client supports stdio MCP.
  2. Verify the package name is @sylphx/pdf-reader-mcp and check the requested version.
  3. Use an absolute PDF path and confirm the client process can read it.
  4. If a scanned document is empty, check the OCR provider and MCP_PDF_OCR_* environment configuration.
  5. When evidence is insufficient, call search_pdf first, then use pdf_evidence with render_page or extract_regions to verify the source region.

Things to try

Once connected, you can ask your AI assistant things like:

  • Read the local PDF at /path/to/report.pdf and extract the table on page 1 with page and bbox citations.
  • Search /path/to/report.pdf for the term "revenue" and return page numbers and bounding boxes for each hit.
  • Run OCR on pages 2-3 of /path/to/scan.pdf and cite the page region for each extracted passage.
  • Inspect the table region on page 4 of /path/to/contract.pdf and report cell count and confidence.

Tools 3

read_pdf read-only
Reads a PDF, automatically selects an extraction route, and returns Markdown, text, tables, chunks, a document map, and related reports.
search_pdf read-only
Searches selectable text and optional OCR text, returning snippets, offsets, pages, bounding boxes, and provenance.
pdf_evidence read-only
Performs inspection, page rendering, region extraction, page OCR, and region analysis for focused PDF verification.

Use cases

Read and summarize digital PDFs.
Search for a term and verify the source using page and bounding-box evidence.
Extract tables, citation chunks, and structured document content.
Process scanned PDFs with OCR and word-box provenance.
Inspect hidden text, suspicious content, and unsafe links.

Supported clients

Claude Desktop
Claude CodePartial support
CursorPartial support
VS CodePartial support

Listed from the project's documentation, not tested by this site.

Overview

PDF Reader MCP is a local-first PDF MCP server that provides an Agent Document Twin with Markdown, structured elements, tables, document maps, citation chunks, page and bounding-box evidence, visual crops, OCR provenance, trust reports, and accessibility reports.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 659 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision 07700c78d9a3 Data synced 2026-10-19 Read the FMRS scoring method