docsagent/docsagent

What it solves

DocsAgent MCP provides a way for AI agents (like Claude Desktop, Cursor, or Cline) to interact with a user's local Zotero research library. It solves the problem of giving LLMs access to thousands of academic PDFs and notes while maintaining strict local privacy and high-speed retrieval, avoiding the need to upload sensitive research data to a cloud service.

How it works

The system uses a three-tier architecture:

1. MCP Shell: A TypeScript or Python wrapper that implements the Model Context Protocol, providing a standardized interface for AI clients to call tools.
2. DocsAgent Core: A resident C++ search engine that runs as a background service. It reads the Zotero database (zotero.sqlite) and storage files directly from the local machine to build a full-text BM25 index.
3. Zotero Local API: Used by the shell to orchestrate write operations (like adding notes or importing items) back into the Zotero library.

Who it’s for

Researchers, academics, and students who use Zotero to manage their literature and want to use AI agents to search, summarize, and organize their papers without compromising data privacy.

Highlights

• High Performance: Average retrieval latency of ~15ms using a native C++ core, even for libraries with over 1,000 PDFs.
• Local-First Privacy: PDFs and database files never leave the local machine.
• Comprehensive Toolset: Includes 8 MCP tools for reading (searching, getting content/metadata, listing the library) and writing (importing items, adding notes, batch modifying tags/collections).
• Write Safety: A three-layer safety gate prevents accidental modifications, requiring explicit configuration and confirmation for write actions.
• Flexible Deployment: Supports both stdio for local clients and Streamable HTTP for remote deployment with RBAC and OAuth 2.0 support.