MCP Server

Beaver includes a built-in MCP (Model Context Protocol) server that lets Claude, ChatGPT, or other AI coding and writing tools access your Zotero library directly — search references, read PDFs and notes, browse collections, and create notes with citations.

What You Can Do

With the MCP server enabled, any compatible AI tool can:

• Search by topic: find papers in your library using semantic (meaning-based) search, the same search that powers Beaver's chat (fulltext search from the beta/pro version is currently not supported)
• Search by metadata: look up specific papers by author, title, or journal
• Read PDFs: extract and read the full text of attachments stored in Zotero
• Read notes: read existing Zotero notes as simplified HTML, including citation tags (Beaver version 0.20 and later)
• Create notes: create new notes from markdown with <citation> tags pointing at items in your library (Beaver version 0.20 and later)
• Get item details: retrieve complete bibliographic metadata (DOI, abstract, authors, etc.)
• Browse collections and tags: explore how your library is organized
• List items: enumerate items in a collection or with a specific tag

This set of tools is designed to make your research library discoverable and searchable by AI tools. If you are looking for a broader set of tools, take a look at Zotero MCP, a dedicated MCP server for Zotero.

Read-Only by Default

Eight of Beaver's nine MCP tools are read-only — they search and read your library but cannot modify or delete anything. The exception is create_note, which can add new notes to your library. It cannot edit or delete existing items.

MCP clients may treat create_note differently from the read tools. For example, asking for approval before running it. Beaver advertises this distinction to clients via standard MCP tool annotations but the actual approval UI is left to the client to implement.

Requirements

• Beaver 0.13 or later installed, with a Beaver account signed in
• For Claude Desktop or other MCP clients that require HTTPS: Node.js installed

Setup

Step 1: Enable the MCP Server

That's it! Your MCP server is running and ready to use with other AI tools.

Step 2: Configure Your AI Tool

Beaver provides two configuration options depending on what your MCP client supports. In Beaver settings under Advanced, you'll find two Copy buttons.

Clients that support HTTP (e.g., Claude Code, Cursor, etc.)

Claude Code, Cursor, Windsurf, and most other MCP clients can connect to Beaver's HTTP endpoint directly — no bridge script or Node.js needed.

The HTTP configuration looks like this:

{
  "mcpServers": {
    "beaver-zotero": {
      "type": "streamable-http",
      "url": "http://localhost:23119/beaver/mcp"
    }
  }
}

Clients that require HTTPS (e.g., Claude Desktop)

Claude Desktop doesn't support plain HTTP connections, so it can't connect to Beaver's local endpoint directly. Beaver provides a small bridge script that handles this automatically.

{
  "preferences": {
    "coworkScheduledTasksEnabled": true,
    "sidebarMode": "chat",
    "coworkWebSearchEnabled": true
  }
}

{
  "preferences": {
    "coworkScheduledTasksEnabled": true,
    "sidebarMode": "chat",
    "coworkWebSearchEnabled": true
  },
  "mcpServers": {
    "beaver-zotero": {
      "command": "node",
      "args": ["/path/to/beaver-mcp-stdio.mjs"]
    }
  }
}