DocScan MCP Server

DocScan is a Model Context Protocol (MCP) server that allows you to scan, index, and search through your local documents using advanced document understanding (via Docling) and local vector search.

Features

• Multi-format Support: Parses PDF, DOCX, PPTX, XLSX, HTML, Markdown, and more.

• Image Support (Performance-first): OCRs standalone images and selectively OCRs weak/empty-text PDF pages.

• Advanced PDF Parsing: Uses Docling for high-fidelity document understanding, including tables and complex layouts.

• Vector Search: Indexes document chunks into a local persistent vector index for semantic search.

• Asynchronous Indexing: Decouples indexing from searching to prevent timeouts on large document sets.

• MCP Integration: Exposes search_docs and reindex_docs tools.

Related MCP server: vectorise-mcp

Setup

Prerequisites

• Python (preferably 3.14)

• uv (recommended for dependency management)

Installation

1. Clone this repository.

2. Install dependencies:

   uv sync

3. Put uv run src/main.py in your MCP config

Example for Copilot:

{
  "mcpServers": {
    "docscan": {
      "type": "stdio",
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/docscan", "run", "src/main.py"]
    }
  }
}

Adding Documents

Place the documents you want to index into the store/ directory. Supported formats include:

• .pdf, .docx, .pptx, .xlsx

• .md, .txt, .html

• .png, .jpg, .jpeg (via OCR)

Usage

Running the Server

You can run the server directly:

uv run src/main.py

Or via fastmcp

fastmcp run src/main.py --transport http --port 8001 --host 0.0.0.0

(change transport and arguments to taste)

Indexing Documents

Before you can search, you must build the index. Use the reindex_docs tool from your MCP client. This only needs to be done once or when you add/change files. Note: Large documents (like some PDFs) may take a few minutes to process.

reindex_docs is incremental: it persists a manifest (vectorstore.db.manifest.json) and only reprocesses files whose size/mtime changed, reuses unchanged chunks and embeddings, and removes deleted files from the index.