You have three hundred papers in Zotero. You have read maybe forty of them properly. You remember roughly nine.
That gap is the reason Zotero MCP is worth twenty minutes of setup. Not because it reads the papers for you, but because it closes the distance between “I know I saved something about this” and actually finding it.
- 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.
- Full-text search is the killer feature, because it reaches inside your PDFs rather than just titles.
- Everything stays local unless you ask the assistant to quote something back to you.
What it changes
A reference manager is very good at storing things and quite bad at reminding you what is in them. Zotero knows you have a paper. It does not know that the paper contains the exact counter-argument you needed last Tuesday.
Connecting it to an assistant changes the questions you can ask from bibliographic to substantive.
Before, you searched for a title you half-remembered. After, you ask which of your saved papers disagree with each other on a specific point, and get an answer drawn from the full text of things you personally chose to keep.
That last part matters more than it sounds. An AI assistant answering from training data is working from a blurry average of the internet. An assistant answering from your library is working from sources you selected, vetted and trust. Same model, completely different quality of answer, because the input changed.
The Model Context Protocol, or MCP, is the open standard that makes this possible. Introduced by Anthropic in late 2024, it gives assistants one common way to reach outside tools. Your Zotero library becomes a source the assistant can query mid-conversation, alongside your files and anything else you have connected.
The servers worth knowing
There is no official Zotero server, and the honest state of play is that this is a community effort discussed mostly on the Zotero forums.
54yyyu/zotero-mcp is the one most people land on and the one I would start with. It handles both local and web API access, and critically it supports full-text search across your PDF attachments rather than only metadata. It also handles annotations and notes, so your own highlights become searchable.
cookjohn/zotero-mcp is a second implementation with a similar goal. Worth knowing about if the first one does not fit your setup.
Both are free and open source. Neither is a product with a support desk, which is the trade you make for something that reads a library you already own.
Local API or web API
This is the decision that determines how the thing behaves day to day, and it is worth getting right first time.
- Reads your complete library, including items never synced.
- Full-text search across attachments actually works.
- Nothing leaves your machine to fetch results.
- Requires Zotero to be running, every time.
- Works whether or not Zotero is open.
- Reaches your library from any machine.
- Only sees what you have synced to Zotero’s servers.
- Needs an API key, and full-text depth is more limited.
For most people the local API is the right answer. Your library is on your machine, you are working at that machine, and full-text search is the feature that makes this worth doing at all.
Enable it in Zotero under Settings, then Advanced, with the checkbox marked “Allow other applications on this computer to communicate with Zotero”. Then leave Zotero running. The most common support question about these servers is someone asking why they get empty results, and the answer is almost always that Zotero is closed.
What to actually ask
Setup guides stop at “it works”. Here is what makes it worth the twenty minutes.
- Find the argument, not the title. “Which of my papers argue against the standard interpretation of [X]?” This is the query a reference manager fundamentally cannot answer, and full-text search plus a model can.
- Surface your own forgotten highlights. “What did I annotate about methodology limitations?” Your highlights are the most valuable and least revisited thing in your library.
- Spot the gap. “Based on my library, what perspective am I missing on this topic?” Genuinely useful before you write, and slightly uncomfortable in a productive way.
- 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 sitting on your disk.
- Triage the backlog. “Which unread papers in my library are most relevant to [current project]?” Turns three hundred saved PDFs back into a reading list.
The pattern is that every one of these is a question you would never open Zotero to answer, because the effort was higher than the payoff. Drop the effort and the questions get asked.
The honest limits
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 it will not tell you the slice is narrow. That is a feature when you want your own sources and a trap when you forget.
Full-text depends on Zotero’s index. If Zotero has not indexed a PDF, or the PDF is a scan without a text layer, the server cannot read inside it. Check Zotero’s own full-text search works before blaming the server.
Large libraries get slow. Several thousand items with full attachments will 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 “far more reliable” is not “verified”. Open the paper before the claim goes in your thesis.
Where ContextBolt fits
I should be straight about this. ContextBolt does not integrate with Zotero, and I am not going to invent a connection that does not exist.
It solves the adjacent half of the same problem. Zotero holds the papers you deliberately filed. ContextBolt holds the other pile, the threads, articles and posts you saved in passing from X, Reddit, LinkedIn and the open web, which is where a surprising amount of current thinking actually lives. It tags and clusters them automatically, and it ships its own MCP server, so Claude can search that pile the same way it searches your Zotero library.
If you are the kind of person who set up Zotero MCP, you probably also have four hundred unsorted bookmarks doing nothing. Same fix, different pile. Pro is $6 a month and there is a free tier to try it on.
The two sit side by side well. Formal sources in Zotero, everything else in ContextBolt, and one assistant able to read both in the same conversation.
Worth doing?
If you use Zotero seriously, yes, and the setup is genuinely twenty minutes.
The thing that surprises people is not the search. It is being asked a question by their own library. You go looking for one paper, and the assistant points out that two of your sources contradict each other on a point you had not noticed. That is the moment the library stops being storage and starts being useful.