Guide · Zotero MCP

Zotero MCP: Connect Your Library to Claude (2026)

Zotero MCP connects your research library to Claude, so you can ask questions of the papers you’ve collected instead of searching for titles. There’s no official Zotero server, but a community one is widely used and takes about twenty minutes to set up.

It runs on the Model Context Protocol, the open standard introduced by Anthropic in late 2024, and it reaches inside your PDFs as well as titles and abstracts.

Quick answer
  • There is no official Zotero server. The community built it, and the Zotero forums are where it gets discussed.
  • 54yyyu/zotero-mcp is the main one. Start there unless you have a reason not to.
  • Two access routes. Local API needs Zotero running, web API needs a synced library.
  • Reads stay local unless you ask the assistant to quote something back to you, or you turn on optional semantic search with a cloud embedding model. The server can also write to your library.

How do you connect Zotero to Claude?

Use the main community server, 54yyyu/zotero-mcp, which reads your local library and supports full-text search across attachments. It needs Python 3.10 or later and Zotero 7 or later.

  1. Install the server. Run uv tool install zotero-mcp-server, or pip install zotero-mcp-server.
  2. Enable the local API in Zotero. Go to Settings, then Advanced, and tick “Allow other applications on this computer to communicate with Zotero”. Leave Zotero running while you use the server.
  3. Add the server to your client. For Claude Desktop, run zotero-mcp setup and it writes the config for you. By hand, add a zotero server with the command zotero-mcp and the environment variable ZOTERO_LOCAL set to true, in claude_desktop_config.json for Claude Desktop, mcp.json for Cursor, or config.toml for Codex. Restart the client afterward.
  4. Test it. Ask “what papers do I have on [your topic]?” If titles and authors come back from your own library, you’re connected. If nothing returns, check that Zotero is running and the local API checkbox is ticked.

Writes are optional. On Zotero 10 or later, run zotero-mcp authorize-local once and choose Always Allow. On older Zotero, add a Zotero API key and library ID so it writes through the web API. In local mode the server reads your library from Zotero’s database file, but anything it writes goes through the running app.

Is there an official Zotero MCP server?

No. Everything available is community-built, discussed mostly on the Zotero forums. Both of these are free and open source, and neither comes with a support desk.

54yyyu/zotero-mcp is the one the community lands on most and the one I’d start with. It handles both local and web API access, supports full-text search across your PDF attachments rather than only metadata, and handles annotations and notes, so your own highlights become searchable.

cookjohn/zotero-mcp takes a different route. It’s a plugin that runs inside Zotero rather than a separate server, which helps if the first one doesn’t fit your setup.

Local API or web API?

Start with the local API. Your library is on your machine, you’re working at that machine, and full-text search is the reason to do this at all.

Local APIWeb API
What it readsYour complete library, including items never syncedOnly what you’ve synced to Zotero’s servers
Zotero open?Yes, writes need the running appNo, it works either way
Full-text searchWorks across attachmentsMore limited
SetupLocal API checkbox in ZoteroA Zotero API key; reaches your library from any machine

Local reads send nothing off your machine to fetch results.

What can you ask your library?

Every one of these is a question you’d never open Zotero to answer, because the effort was higher than the payoff.

  • Find the argument, not the title: “Which of my papers argue against the standard interpretation of [X]?” A reference manager can’t answer this, and full-text search plus a model can.
  • Surface your forgotten highlights: “What did I annotate about methodology limitations?”
  • Spot the gap: “Based on my library, what perspective am I missing on this topic?”
  • Draft with real citations: “Summarize what my sources say about [X] and cite which paper each claim came from.” The citations are checkable, because they point at things on your disk.
  • Triage the backlog: “Which unread papers in my library are most relevant to [current project]?”

An assistant answering from training data works from a blurry average of the internet. One answering from your library works from sources you chose, so the same model gives a better answer.

What are the limits of Zotero MCP?

Your library is not a search engine. If you saved forty papers on a topic, the assistant answers from forty papers. It will sound authoritative about a narrow slice and gives no hint that the slice is narrow.

Full-text depends on Zotero’s index. If Zotero hasn’t indexed a PDF, or the PDF is a scan without a text layer, the server can’t read inside it. Check that Zotero’s own full-text search works before blaming the server.

Large libraries get slow. Several thousand items with full attachments make some queries take a while. Narrow by collection or tag when you can.

Citations still need checking. A model quoting your library is far more reliable than one quoting from memory, but that isn’t verified. Open the paper before the claim goes in your thesis.

Writes are possible. The main server can add papers, annotate PDFs, manage collections and tags, and merge duplicates, so use your client’s confirmation prompts. If your library holds confidential or unpublished work, treat anything you ask about it the way you’d treat pasting it into a chat window.

Where does ContextBolt fit?

Zotero MCP reaches the papers you deliberately filed. It doesn’t reach the threads, articles and posts you saved in passing from X, Reddit, LinkedIn and the open web. ContextBolt holds that pile. It’s a Chrome extension that captures those saves as you make them and tags and clusters each one automatically. On Pro it ships a personal MCP endpoint, so the same Claude that searches your Zotero library can search your saved web by meaning in the same conversation.

The two don’t integrate, and don’t need to. Zotero stays the home for formal sources, and one assistant reads both. Basic is free up to 150 bookmarks, with no account and no card. Pro is $6 a month, removes the cap, and adds the MCP endpoint.

Where next

Frequently asked questions

Something else on your mind? Get in touch.

What is Zotero MCP?

It is a small server that lets an AI assistant read your Zotero library through the Model Context Protocol. Once connected, you can ask Claude or ChatGPT about the papers you have collected, search across their full text, pull citations, and have the assistant reason about your own sources rather than whatever it half-remembers from training.

Is there an official Zotero MCP server?

No. Zotero has not shipped one. Everything available is community-built, and the discussion happens on the official Zotero forums rather than in official documentation. The most widely used option is 54yyyu/zotero-mcp.

Does Zotero MCP read the full text of my PDFs?

The better servers do. If your PDFs are stored as attachments and Zotero has indexed them, full-text search reaches inside them rather than only matching titles and abstracts. That is the difference between asking what a paper is called and asking what it actually argues.

Do I need Zotero open for the MCP server to work?

Keep it open for local access. The local setup relies on Zotero's local API setting, and anything the server writes to your library goes through the running app. Web API access works without Zotero running, but only sees items you have synced to Zotero's servers.

Is it safe to connect Zotero to an AI assistant?

Local access stays on your machine, but the server is not read-only. The main community server can also add papers, annotate PDFs, manage collections and tags, and merge duplicates, so use the confirmation prompts in your AI client. Nothing is uploaded unless you ask the assistant to quote something, in which case that text goes to the model like any other message you send. If you switch on the optional semantic search with an OpenAI or Gemini embedding model, the text it indexes goes to that provider too. If your library contains confidential or unpublished work, treat anything you ask about it the same way you treat pasting it into a chat window.