File Analysis MCP

A personal MCP server that reads unstructured documents (PDF, DOCX, PPTX, SVG, PNG, JPG) in a specific folder and extracts their structure/content. Summarization is performed by the LLM in Claude Code/Codex that calls this server, not by this server itself. This server only handles text/image extraction and folder structure analysis.

Prerequisites

1. Python

python and pip must work properly. If python on Windows shows the Microsoft Store installation prompt, check the following:

1. Install Python 3.11 or later (64-bit) from python.org, and check "Add python.exe to PATH" during installation.

2. If python still points to the Store stub: go to Settings → Apps → Advanced app settings → App execution aliases and turn off the python.exe/python3.exe entries for App Installer.

2. Tesseract-OCR (required for PNG OCR, optional)

Install it via the UB-Mannheim Tesseract installer, then set the TESSERACT_CMD value in .mcp.json to the actual installation path. The server works fine even without it; when analyzing PNGs, the image itself (multimodal) is still returned and only the OCR text is replaced with an "unavailable" message.

Related MCP server: ReadPDFx - OCR PDF MCP Server

Installation

cd C:\Users\20229\Desktop\MCP
python -m venv .venv
.venv\Scripts\pip install -r requirements.txt

Verification

# 1. 샘플 파일 생성 (pdf/docx/pptx/svg/png/txt 각 1개)
.venv\Scripts\python scripts\generate_fixtures.py

# 2. MCP 없이 extractor만 직접 테스트
.venv\Scripts\python scripts\smoke_test.py

Registering with Claude Code

The .mcp.json in this folder already registers the server at the project scope. After checking that the command path matches the actual .venv\Scripts\python.exe location, restart Claude Code or use the /mcp command to confirm the file-analysis server is in the connected state.

Provided Tools

• scan_folder(folder_path, max_depth=3, max_files=200): Returns only the folder tree and file counts by extension (does not read contents).

• analyze_file(file_path, max_chars=None): Extracts a single file according to its extension.

  • pdf/docx/pptx/svg → text/structure information

  • png/jpg/jpeg → image (multimodal) + OCR text

Limitations

• SVG only parses the XML structure (text nodes, shape count, viewBox). For pure vector graphics without text, it's impossible to tell what is being drawn visually — convert to PNG and re-analyze if needed.

• Extraction is capped at limits such as PDF 50 pages/50,000 characters, PPTX 100 slides/50,000 characters, and PNG OCR 20,000 characters. When exceeded, the result includes truncated: true and a notice message.