Zotero group libraries with Claude for systematic review teams

A systematic review team keeps its references in one place: a Zotero group library, often with five to ten thousand records, shared between a PI, postdocs, students and a librarian. This page shows how to put that library inside Claude so the whole team can search it by meaning, pull quoted passages with page numbers, tag screening decisions, and build the bibliography, while Zotero stays the system of record.

What Claude can do with a shared library

With Zoteus connected to the group, every member can ask Claude to:

• Search the group by keyword or by meaning. "Which of our included studies measured adherence at twelve months?" runs against titles, abstracts, creators, tags, your notes and your PDF annotations, and optionally available PDF body text within configured indexing caps.
• Quote the evidence with a locator. zotero_get_fulltext returns the passage with its page number and the nearest heading, so an extraction table can cite the page a number came from.
• Tag and file decisions. Add or remove tags (include, exclude-wrong-population, full-text-pending) and move items between collections, so the screening state lives in Zotero where everyone sees it.
• Audit the tagging. zotero_tag_audit checks the library against a controlled vocabulary you supply and reports off-taxonomy tags and items missing a required tag, per collection if you want.
• Add what is missing. Add a paper to the group by DOI or arXiv id, and attach the PDF from a URL.
• Chase citations. zotero_scholar looks up a paper's references and citing works through OpenAlex (Crossref as fallback) and flags which of them the library already holds, which is backward and forward snowballing in one call.
• Build the bibliography in any CSL style with citeproc-js, or export RIS, CSV, BibTeX or CSL JSON for another tool.

What it does not do: it does not replace dual independent screening, and it does not decide anything on its own. A person asks, reads the answer, and acts. Writes go to Zotero with version checks, "remove" means the trash (reversible), and permanent deletion is switched off unless the operator turns it on and Claude confirms on every call.

Individual or Lab?

The plans sell hosting, seats and support, not group-library functionality. Individual covers one researcher, including libraries their Zotero permissions allow. Lab covers up to ten people with assisted setup, priority email support, invoicing and purchase orders. It does not unlock combined semantic search across libraries. Compare the plans and prerequisites.

Two ways to run it

A local connector can read local files, but retrieved excerpts still go to the chosen AI service. Local embeddings do not make Claude or ChatGPT run locally. If institutional policy requires data to stay on its machines, assess the entire client, connector, embedding and Zotero data path. Hosted access suits teams that prefer a maintained server and whose data requirements permit it. Details on both are on the pricing page.

Setting it up

1. Have a group library in Zotero

If you do not have one yet, create it at zotero.org/groups, choose the membership setting that fits (private is usual for a review), and invite the team. Members have to accept the invitation and sync at least once. Zotero's own group documentation covers roles: admins can change settings and membership, members can edit items, and a library can be set so that only admins edit.

2. Give each member access

Self-hosted. Each member creates an API key at zotero.org/settings/keys: tick Allow library access, and under the group permissions give the review group Read/Write (or set the default group permission to Read/Write). They then paste that key into their Zoteus install; the Connect Claude to Zotero page has the exact steps for each client. On Zotero 10 and newer, a group that the desktop app already holds can be read without any key while the app is running; writes to a group always need the key.

Hosted. Before checkout, the billing owner arranges member access with support. Each member uses their assigned access, adds https://mcp.zoteus.com/mcp as a custom connector in claude.ai or as a plugin in ChatGPT, and approves Zoteus on zotero.org. No API key is created by hand.

3. Point Claude at the group

Ask Which Zotero group libraries can you see? Claude calls zotero_groups, which lists each group with its id, item count and your edit permission. From then on, name the group in your request ("in the Cochrane update library, ...") and Claude passes its library_id to the tools.

4. Build the search index once

First check search readiness, including the selected library, active embeddings, item cap and PDF coverage. Then ask to build the index for that group's numeric ID. Direct keyword lookup works before the index exists.

Personal and group indexes side by side: each library gets its own search index in the same data directory (per authenticated user when hosted), so indexing the group does not touch your personal index. Ask Claude to search both at once; each hit is labelled with the library it came from.

Default indexing stops at 5,000 items. PDF bodies are opt-in and default to 40,000 characters per item; only available extracted text is included. Those limits do not prevent reading a later page directly from an accessible PDF. Check one verified passage before a large build.

5. Work

A few prompts that review teams start with:

• List everything in the "screening" collection tagged full-text-pending and, for each, quote the sentence that states the primary outcome, with the page.
• Tag every item in this collection whose abstract mentions a paediatric population with population-children. Show me the list first.
• Audit the tags in the group against this vocabulary: [paste the vocabulary]. Which items are missing a required tag?
• Look up the references of 10.1136/bmj.n71 and tell me which ones are not yet in the group library.
• Format the included studies as a bibliography in Vancouver style.

Zoteus also ships MCP Prompts for common workflows, including a citation audit that checks a draft's citations against the library; see Prompts.

Data handling, for the ethics form

• Self-hosted: tool results, including retrieved passages and notes, are sent to the chosen AI client/service. Zotero sync/cloud calls and optional external embedding or scholarly services have their own data paths. Local embeddings keep embedding computation on that host only. Optional usage logging is off by default and records tool outcomes without query text; the operator controls it.
• Hosted: Zoteus stores each member's Zotero token encrypted at rest (AES-256-GCM) and uses it only for that member's requests; operational logs, with secrets redacted, are kept for up to 30 days. The hosted instance currently runs on Google Cloud in the United States. A Data Processing Addendum (GDPR Article 28) is available on request. The full privacy policy has the sub-processor list.
• Seats: choose the plan that covers the people using it. Arrange member access with support during setup; a license key can be bound to a Zotero account, so do not assume the same key can be shared among everyone.

The pilot

For a review team the setup, not the software, is the part that costs an afternoon, so it is done with you. A pilot is:

1. A 15-minute call with the maintainer, on which your group library is connected and the first index is started. Bring one known paper and one research question to check PDF availability and retrieve a verifiable passage. Agree on a task to evaluate during the pilot.
2. Thirty days of the Lab plan at no charge, for the whole team. No card.
3. At the end, either the Lab plan (€99 a month, or €990 a year, invoices and purchase orders are available) or nothing happens.

Book the setup call or email support@zoteus.com. If your group is larger than ten, the Department plan covers up to fifty people and comes with an onboarding session; write first.