PDF Text Extractor MCP

PDF Text Extractor MCP

A local STDIO MCP server exposing exactly one tool, extract_pdf_text. It reads embedded text with PDF.js, falls back to English/German OCR for scanned pages, and returns page-level paragraphs in document order. No React app, HTTP server, API key, or PDF upload is needed.

Requires Node.js 22.19 or newer and npm (node and npx on PATH).

Run the published package

Version 0.1.1 is available on npm. Launch it without cloning the repository or starting the React app:

npx -y pdf-text-extractor-mcp@0.1.1

The process waits for an MCP client on stdin. There is no interactive prompt. The client normally launches this command itself.

VS Code configuration

Put this in .vscode/mcp.json, then start the server from VS Code's MCP controls and enable the tool in agent chat:

{

"servers": { "pdf-text-extractor": { "type": "stdio", "command": "npx", "args": ["-y", "pdf-text-extractor-mcp@0.1.1"] } } }

For a local development build, use "command": "node" and "args": ["/absolute/path/to/pdf-text-extractor/mcp/dist/index.js"] after building locally. Other clients may call the top-level key mcpServers instead of servers; follow that client's configuration format.

Ask the client to extract a local PDF, or invoke extract_pdf_text with:

{ "file_path": "/absolute/path/to/contract.pdf" }

In JSON on Windows use an absolute path such as C:\\Users\\you\\contract.pdf. The path is on the machine running the MCP server.

Result contract

Successful calls return isError: false, a JSON text content block, and the same object in structuredContent:

{

"success": true, "text": "First page text\n\nSecond page text", "page_count": 2, "pages_without_text": [], "pages": [ { "page_number": 1, "method": "embedded", "text": "First page text", "paragraphs": ["First page text"], "warnings": [] }, { "page_number": 2, "method": "ocr", "text": "Second page text", "paragraphs": ["Second page text"], "warnings": [] } ], "warnings": [] }

File or extraction errors return isError: true with { "success": false, "text": "", "error": "PDF file was not found." }. Missing or incorrectly typed arguments are rejected by the SDK with an MCP tool error before extraction runs.

The original text, page_count, and pages_without_text fields remain available. text joins nonempty pages with blank lines; each page's text joins its paragraphs with blank lines and preserves line breaks within paragraphs. page_number and pages_without_text use one-based page numbers. method is embedded, ocr, or none when no text could be extracted and OCR failed. An OCR attempt that completes but finds no text reports ocr with a warning. Top-level warnings repeat page warnings with their page number. success: true means the PDF was processed; consult warnings for partial results.

Only regular .pdf files with a PDF header are accepted, up to 25 MiB. Missing, unreadable, invalid, oversized, and password-protected files produce errors. The server reads files without modifying them, executes no shell commands, and makes no network requests for extraction. Diagnostics go to stderr; stdout is reserved for MCP messages. PDF contents are never logged.

Limitations

Pages with no embedded body text, including header/footer-only pages, use OCR. Pages with fewer than 40 body-text characters and an image also use OCR. Successful OCR replaces the page's embedded text to avoid duplicate headers. If OCR fails or returns no text, any embedded text is retained and a warning is returned. Pages still without text appear in pages_without_text.

English and German language models are bundled at build time. Extraction needs no network connection, model download, or writable model cache. The OCR worker is started only when needed, reused across the document, and terminated afterward. OCR input is rendered at up to 2.5x PDF scale, capped at 8 million pixels per page. OCR recognition, including engine startup but excluding PDF rendering, has a 60-second page limit. If the worker fails or times out, it is terminated; remaining pages retain their embedded text with warnings.

Embedded text is grouped into rows and paragraphs by position. Large horizontal gaps are represented as | for basic table readability. OCR uses Tesseract's paragraph detection and word positions. These are readability heuristics, not semantic table extraction or complex layout analysis. Multi-column, rotated, handwritten, or poor-quality scanned pages may have incorrect order or text. Pages with substantial embedded text over a scanned body may not trigger OCR. Low OCR confidence generates a warning; higher confidence is not a correctness guarantee. The size limit bounds input bytes, not PDF decompression or processing time. The server can read any PDF accessible to its OS user; there is no directory sandbox. Text is returned to the MCP client, whose own data handling settings apply.

Development

From the repository root:

npm ci --prefix mcp

npm run build --prefix mcp node mcp/dist/index.js

The CLI is a separate package so installing it does not install React. It includes Tesseract.js and a native Node canvas dependency. Use a platform supported by @napi-rs/canvas and allow npm to install its platform binary (do not omit optional dependencies). Development model packages are copied into dist/tessdata/ by the build; consumers do not need development dependencies. Its own lockfile provides a reproducible build. The root app commands are unchanged.

The browser service and this CLI use the same runtime-independent extraction core in src/shared/: page processing, embedded-text reading, scan detection, OCR fallback, paragraphs, table spacing and warnings. Browser and Node adapters retain their own file access, PDF.js setup, rendering limits and OCR worker management. The browser's existing result interface and this tool's JSON contract are preserved. The build cleans generated output and includes compiled dist/shared/ modules in the tarball; runtime consumers need no shared source files from the repository.

Test with MCP Inspector

Test the published package without cloning or building:

npx -y @modelcontextprotocol/inspector@2.13.0 --cli npx -y pdf-text-extractor-mcp@0.1.1 --method tools/list

npx -y @modelcontextprotocol/inspector@2.13.0 --cli npx -y pdf-text-extractor-mcp@0.1.1 --method tools/call --tool-name extract_pdf_text --tool-arg "file_path=/absolute/path/to/contract.pdf"

For the browser Inspector:

npx -y @modelcontextprotocol/inspector@2.13.0 npx -y pdf-text-extractor-mcp@0.1.1

Click Connect, open Tools, and invoke extract_pdf_text with a local PDF path. Inspector is development tooling and is not a runtime dependency of this package.

To inspect a local development build, run from the repository root after building:

npx -y @modelcontextprotocol/inspector@2.13.0 --cli node mcp/dist/index.js --method tools/list

npx -y @modelcontextprotocol/inspector@2.13.0 --cli node mcp/dist/index.js --method tools/call --tool-name extract_pdf_text --tool-arg "file_path=/absolute/path/to/contract.pdf"

Automated tests and packaging

npm test --prefix mcp

cd mcp npm pack

Inspect the tarball with tar -tzf pdf-text-extractor-mcp-0.1.1.tgz. It contains compiled JavaScript (including the OCR worker), both compressed language models, third-party notices, this README and package metadata. Install it in a separate directory to verify the executable without the source checkout:

mkdir /tmp/pdf-mcp-install

cd /tmp/pdf-mcp-install npm install /absolute/path/to/mcp/pdf-text-extractor-mcp-0.1.1.tgz ./node_modules/.bin/pdf-text-extractor-mcp

To run the integration test against that installed executable from the repo root:

MCP_TEST_CLI=/tmp/pdf-mcp-install/node_modules/.bin/pdf-text-extractor-mcp npm test --prefix mcp

Publish a new version

Published versions cannot be overwritten; increment the version for the next release and update pinned version examples in both READMEs. Project code is licensed under MIT; the package includes LICENSE with copyright © 2026 Prashant Bhatia. Bundled OCR models retain their Apache-2.0 license and attribution in THIRD_PARTY_NOTICES.md and licenses/Apache-2.0.txt.

From mcp/:

npm login --auth-type=web

npm version patch --no-git-tag-version npm test npm pack npm publish --dry-run --access public npm publish --access public --auth-type=web

Inspect and test the packed artifact before publishing, as described above. Publishing requires access to the maintainer's npm account with 2FA enabled; complete any browser approval prompted by npm. Commit the version, lockfile and documentation changes for the release. The build and tests do not publish automatically. Publish from mcp/, since the root React package remains private.

Publish to the MCP Registry

The registry identifier is io.github.Prashant2204/pdf-text-extractor-mcp. The mcpName in package.json must match name in server.json. Keep the package version, server version, and npm package version in server.json aligned. Publish the npm package first so the registry can verify its metadata.

Install the official publisher with brew install mcp-publisher, then from mcp/:

mcp-publisher validate

mcp-publisher login github mcp-publisher publish

Complete GitHub device authorization as prompted. Verify the published entry with:

curl "https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.Prashant2204/pdf-text-extractor-mcp"

The registry lists metadata; npm distributes the local STDIO server. See the official publishing guide.

The integration test runs real English/German OCR on generated image-only PDF pages, mixed header/scan pages, blank pages, and embedded-text pages through STDIO. Other tests cover ordering, table gaps, OCR timeouts, and partial results when image inspection or OCR assets fail. The same integration test can target the installed executable as above. See THIRD_PARTY_NOTICES.md for bundled model attribution.

Official references

• SDK first server
• SDK STDIO transport
• MCP Inspector
• VS Code MCP servers

Implementation verified against stable @modelcontextprotocol/server 2.4.0, using McpServer.registerTool() and serveStdio(). Browser and MCP packages both use PDF.js 6.5.438.