PDF Reader MCP is now AnyMD, a local document-to-Markdown server that gives assistants page-addressable PDF text. It helps when you need to find a passage, answer from it and return to the original page. Older package names still install, but old read_pdf examples no longer match the current interface. For a few pasted lines you do not need a server; repeated work across PDF pages is the stronger use case.
Install and connect
Use Node.js 18 or newer and PowerShell in a fresh directory. The compatibility package below launches AnyMD of the same pinned version:
npm.cmd install --save-exact --no-audit --no-fund @sylphx/pdf-reader-mcp@8.5.11
New-Item -ItemType Directory -Force documents | Out-Null
$node = (Get-Command node.exe).Source
$entry = (Resolve-Path .\node_modules\@sylphx\pdf-reader-mcp\bin\pdf-reader-mcp.js).Path
$docs = (Resolve-Path .\documents).Path
$cache = Join-Path (Get-Location) 'anymd-cache'
Claude Code:
claude.cmd mcp add pdf --scope project --env "ANYMD_CACHE_DIR=$cache" -- $node $entry "--allow-dir=$docs"
claude.cmd mcp list
Codex:
codex mcp add pdf --env "ANYMD_CACHE_DIR=$cache" -- $node $entry "--allow-dir=$docs"
codex mcp list
Claude's health check should report Connected. Start a fresh assistant session before requesting document work. This setup allows only the documents folder and places the cache in the project. It avoids the broad setup command that changes several client configurations at once.
For Cursor, follow the upstream example, placing the same command, entry point and allow-dir argument under mcpServers in .cursor/mcp.json. npm supplies the Windows native binary; a Rust build is unnecessary.
Make a reproducible two-page PDF
Use your own text PDF, or create this tiny one with uv:
uv venv --python 3.12 .venv
uv pip install --python .\.venv\Scripts\python.exe reportlab==5.1.0
Save as make_pdf.py and run .\.venv\Scripts\python.exe make_pdf.py:
from reportlab.pdfgen import canvas
c = canvas.Canvas('documents/sales.pdf')
c.drawString(72, 740, 'Taipei sales: 200 TWD')
c.showPage()
c.drawString(72, 740, 'Kaohsiung sales: 50 TWD')
c.save()
Three useful prompts
Read one page before asking a question
Use only PDF MCP read. Read the full path to documents/sales.pdf,
with pages="2", ocr=false and images="none". Keep the returned page
marker. Do not scan the directory or read unrelated files.
The response should contain <!-- page 2 --> and Kaohsiung sales: 50 TWD. The 200 on page one is not this page's amount. Selecting pages first keeps unrelated material out of the conversation.
Reconcile information across pages
Read the same PDF with pages="1-2", ocr=false and images="none".
List each city's amount beside its original page number.
Identify missing information instead of filling it in.
In this file, page one contains Taipei/200 and page two Kaohsiung/50. The markers help locate evidence in the original document; they do not prove that every visual element was extracted.
Search and answer with attribution
Use search with query Taipei, mode="literal" and sources containing
only the full path to sales.pdf. Show the matching page and text,
then answer the question: what were Taipei's sales?
Our result was one hit on p.1 containing Taipei sales: 200 TWD. Literal mode makes the requested match explicit instead of retrieving merely related passages.
Troubleshooting
read_pdfis missing: 8.5.11 exposesoutline,read,searchandinspect. Ask the assistant to inspect the current tool schemas. This walkthrough uses onlyreadandsearch.- A scan warning appears despite readable text: our one-line pages triggered
These pages have little or no selectable textwhile both lines were correctly returned. Check the extracted content before installing OCR models. Sparse text is different from an image-only page. - The native executable is missing: check whether optional npm dependencies were omitted. Windows x64 uses
@sylphx/anymd-win32-x64-msvc; reinstall the matching platform package instead of trying to run a Linux binary. - The path is denied or cannot be found: keep the source inside the allowed directory and use its full path. The assistant's working directory and the server's allowed directory need not be the same.
Free local use and limitations
The local core is MIT licensed and reads/searches text PDFs without an account or API key. The separate hosted service is metered and some Pro operations are paid. AI-client usage is separate from the local parser.
Scans require an OCR route and its software or models. Test tables, multiple columns, images and extraction coverage on your actual documents; this fixture demonstrates page attribution and text only. For a programmable page-extraction workflow, see PyMuPDF4LLM. For broader office-document conversion, compare MarkItDown.
FAQ
Does the rename require a new account?
No account is needed for local use. Version 8.5.11 of the old npm name is a compatibility entry point to the same-version AnyMD program. New installations can use the upstream AnyMD name; existing users should check their installed version and update tool names in saved prompts.
Can Claude Code read only selected pages?
The MCP read tool accepts page selections such as 2 or 1-5,8. Ask for original page markers and supporting text in the answer. We exercised the same MCP interface through Codex for one-page and two-page reads.
Does a search hit prove complete extraction?
No. It shows that the returned passage matches the query. Scanned content, omitted pages or unusual layouts need further checks. Reconcile page count and selected range against the source, then inspect important content directly.
What we tested
2026-11-16: Windows, Node 24, compatibility package and AnyMD 8.5.11, ReportLab 5.1.0. Codex performed real page-two selection, two-page reading and literal search; Claude Code registration reported Connected. OCR, paid features, complex tables and Cursor were not tested. References: official README and migration guide, accessed the same day.