Choosing and Safely Evaluating a Zotero MCP Server

This conceptual guide helps you choose and safely evaluate a specific Zotero MCP server before connecting it to a research library. Zotero does not define one universal MCP server or configuration: projects differ in library access, permissions, data flow, and attachment-text support. The goal is to verify the selected server against a disposable collection, not to install an MCP service from a one-size-fits-all recipe.

This is a server-selection and safety checklist, not a universal installation walkthrough. Because server projects and their capabilities change, no single server command or configuration is prescribed here.

Decide what you need the connection to do

Start with one task, such as locating records in a test collection and returning their titles, authors, year, and stable Zotero item link. Keep the first test read-only. If you need PDF text, check separately that the chosen server supports it and how it obtains that text. A record search alone does not establish full-text access.

Evaluate a server before configuring it

1. Inspect the project

   Open the project's current README and release information. Confirm its package or repository identity, supported Zotero access method, prerequisites, installation command, and whether it is maintained. The package name in a blog post is not enough to identify a server reliably.

2. Review permissions and data flow

   List the tools the server exposes. Check whether they can create, edit, or delete records, and whether requests or attachment contents leave your computer. Use a dedicated test collection and the narrowest available permissions. Keep API credentials out of prompts, source control, and shared configuration files.

3. Configure the client from verified instructions

   For local command-line servers, Cursor's current configuration uses an mcpServers object with a command and arguments; its documentation also describes server type and environment fields. Copy the exact command and arguments from the server project, then check that the client can start it. Do not substitute a guessed package or wrapper command.

4. Prepare a disposable test collection

   In Zotero desktop, create a collection named MCP-test-YYYYMMDD beneath the specific library you intend to test (for example, My Library or a test group library). Add one record with a title and author you can verify; a harmless manual placeholder is sufficient for checking retrieval, but it is not a scholarly reference. Zotero's collection guide explains creating collections and adding items. Before querying, identify the account or local library configured by the server and confirm in its own documentation whether it can be restricted to this collection; a collection name in a prompt is not an access-control boundary. If the server cannot be scoped, use only a disposable library/account with no sensitive records.

5. Test with the known record

   Ask the client to find the exact test title and return identifying metadata plus its collection name. Compare every returned field with Zotero, verify that the result belongs to the intended library and test collection, then check that the library remains unchanged.

Use citation keys carefully

Better BibTeX provides citation-key generation and export features for text-based writing workflows. It is not a universal prerequisite for every Zotero MCP server. Zotero 8 also has a native citation-key field; Better BibTeX's current documentation describes migration and compatibility changes around that field. Follow the requirements of the server and your citation workflow rather than installing the add-on solely because an MCP tutorial says it is mandatory.

A citation key is a convenient identifier for a citation workflow, not proof that a reference is correct. Check the underlying Zotero item and its metadata. Before submitting a manuscript, verify the citation style and bibliographic details against the original source.

Ask for bounded, auditable results

Give the assistant a defined collection and a small, answerable request. For example:

In my Zotero test collection, find the record with this exact title: [title].
Return its Zotero item link, title, authors, year, and whether an attachment is listed.
If you cannot find an exact match, say so. Do not edit the record or summarize a paper.

This request separates retrieval from interpretation. Once the metadata matches, you can ask for attachment text only if the server documentation confirms that capability. Treat an AI-generated summary as a pointer back to the paper, not as a substitute for checking the methods and results yourself.

What a local citation-key inventory can and cannot do

The primary R route below uses base R and the same regular expression as the optional Python route. Both extract strings shaped like @key followed by a four-digit year from a Markdown file. They do not parse every citation syntax, resolve keys against Zotero, confirm references, or generate a bibliography. This is a quick inventory aid: it can miss valid keys or match text that is not a citation. The input is synthetic and the example key is invented.

Create the same small fixture in a working directory before running either example:

cat > sample.md <<'EOF'
Background: an illustrative citation @rivera2021.
This sentence has no citation key.
EOF

R · base R (primary route)

text <- paste(readLines("sample.md", warn = FALSE, encoding = "UTF-8"), collapse = "\n")
matches <- gregexpr("@([\\w-]+?\\d{4}[\\w-]*)", text, perl = TRUE)
full_keys <- regmatches(text, matches)[[1]]
keys <- sort(unique(sub("^@", "", full_keys)))
if (identical(full_keys, character(0))) keys <- character(0)
cat("Found", length(keys), "candidate citation keys:\n")
for (key in keys) cat("  @", key, "\n", sep = "")

Save as find_keys.R and run Rscript find_keys.R with sample.md in the same directory. Dependencies: R with base packages only. Expected output is Found 1 candidate citation keys: followed by @rivera2021. With no matching string, it prints Found 0 candidate citation keys: and no key lines. A missing or unreadable input produces an error instead of an empty result. The fixture demonstrates only this regular expression; review candidates against Zotero or your bibliography tool.

Python · optional alternative

import re
from pathlib import Path

text = Path("sample.md").read_text(encoding="utf-8")
keys = sorted(set(re.findall(r"@([\w-]+?\d{4}[\w-]*)", text)))
print(f"Found {len(keys)} candidate citation keys:")
for key in keys:
    print(f"  @{key}")

Save as find_keys.py and run python3 find_keys.py in the same directory. Dependencies: Python 3 standard library only. Given the same sample.md, it should print the same candidate list as the R script. The implementations inventory matching text; neither one confirms that a key is valid or that a reference exists.

Troubleshoot by observing the failure

• The client does not show the server: validate the JSON syntax, command path, and arguments against the current client and server documentation. Check the client's MCP logs for its startup error.
• The server starts but returns no records: confirm the account or local library selected by that server, then repeat a search with a known exact title in the test collection.
• Metadata appears but PDF text does not: determine whether the server advertises full-text retrieval and whether it can access that attachment. Zotero metadata access does not by itself imply attachment access.
• A query changes library contents: stop the server, inspect its write-capable tools and permissions, and restore from a known backup if needed. Do not continue testing against your working library.
• The response looks plausible but differs from Zotero: treat it as a retrieval or interpretation error. Use the Zotero item as the source of record and do not cite unverified details.

Further reading

• MCP overview explains the protocol's role in connecting AI applications with external systems.
• Cursor MCP documentation describes current local server configuration fields and configuration locations.
• Zotero's developer documentation distinguishes local-client APIs and online library access.
• Better BibTeX documentation describes citation-key features and current Zotero compatibility notes.