[Windows] MCP tool interface cannot recognize existing local PDF files

[iiishop]

Environment: Windows 10, Python 3.11, poetry installation, latest mcp-document-reader.

Problem Description:

• In the poetry virtual environment, all path formats (absolute, relative, forward slash, backslash) can be correctly recognized by Python's Path().exists(), and the file does exist.
• However, when calling the MCP tool interface of mcp-document-reader (such as get_pdf_metadata), no matter which path is passed, it always reports "The file ... does not exist."
• File permissions, cwd, and path formats have all been checked and are correct. Directly calling PDFReader in readers.py with Python works fine.

Reproduction Steps:

1. Place the PDF file (e.g., 1.pdf) in the cwd directory of mcp-document-reader.
2. Enter the poetry shell, and manually test Path('1.pdf').exists(), which returns True.
3. Use the MCP tool interface get_pdf_metadata to read the same path, and it always reports the file does not exist.

Expected Result:

• The MCP tool interface should be able to recognize local files just like Python code does.

Additional Info:

• This issue always occurs on Windows. Not tested on Linux.
• Detailed troubleshooting process and test code are attached below.

Appendix:

from pathlib import Path
print(Path("D:/mcp/mcp-document-reader/1.pdf").exists())  # True
print(Path("D:\\mcp\\mcp-document-reader\\1.pdf").exists())  # True
print(Path("1.pdf").exists())  # True
print(Path("./1.pdf").exists())  # True
print(Path(".\\1.pdf").exists())  # True

Thank you for your attention! Please let me know if you need more information.

[iiishop]

Update: After further investigation, the root cause appears to be specific to the MCP tool interface, not the underlying PDF reading logic.

I created a standalone test script using the PDFReader class directly (with the same PDF file and all dependencies installed), and it works perfectly: metadata and content are extracted as expected. This confirms that the file path, permissions, and Python environment are all correct, and the PDF can be read without issue in a direct Python context.

However, when using the MCP tool interface (e.g., get_pdf_metadata), the same file path still results in 'The file ... does not exist.'

Summary:

• The PDFReader class works fine when called directly in Python.
• The MCP tool interface still cannot access the file, even though it exists and is readable.

This suggests the problem is likely with the MCP tool's path handling, environment, or stdio communication, not with the PDF reading implementation itself.

If you need any logs, test scripts, or further details, please let me know. Thank you!