Zotero + AI Integrations That Actually Work in 2026

ARM2 meta-analysis: 200 papers, deduped and tagged in 2 days. Not because I found a magic plugin — because I stopped chasing magic plugins and built a lean stack around what Zotero actually does well. If you've browsed the Zotero plugin ecosystem looking for AI, most of what you'll find is demo-quality, last-committed-in-2022 code. Here's what survives a real workflow.

The Problem With Zotero AI Plugins

The promise is always the same: AI reads your library, surfaces connections, writes synthesis. The reality: most plugins are abandoned, require API keys they don't document properly, or fail silently when your library exceeds a few hundred items. I've spent 3 hours on a plugin that promised "semantic search across your Zotero library" and timed out on anything over 100 items.

What Zotero does natively, and does well: clean metadata export and PDF organization. Everything useful is built on those two things.

Better BibTeX: The Bridge That Actually Matters

Better BibTeX is the integration most researchers underestimate. Its job is keeping citation keys stable across your library — which matters the instant you paste references into a Claude Project and want to cross-reference them back to your source records.

Setup: install the plugin, set a citation key pattern (authEtAl_year_title2words is what I use), and enable auto-export to a .bib file in your project folder. Paste that .bib into a Claude Project system prompt. Claude can now identify every paper by citation key when you ask it to organize references in a draft — and you can trace every output claim back to a Zotero record without manual reconciliation.

One failure mode to know: auto-export can lag on a large open library during an active drafting session. Export manually before any ARM-scale writing session.

PDF Organization: Zotero 7 vs ZotFile

If you're on Zotero 7, the native PDF reader and linked-files manager has replaced most of what ZotFile did. I keep ZotFile installed for one thing only: batch-renaming legacy imports with non-standard filenames. The native organizer handles new imports cleanly but won't retroactively fix a library full of files named document_scan_2021.pdf.

For AI workflows, the practical point isn't which organizer you use. It's that consistently named, extractable PDFs are a prerequisite for anything downstream — including running CiteCheck on a completed draft.

CiteCheck: Verification on the Export

CiteCheck isn't a Zotero plugin. It's a CLI that runs on your exported manuscript — but it's the verification layer that closes the loop on the Zotero workflow.

After drafting with Claude (using the Better BibTeX export as reference context), export the draft as plain text and run:

pip install citecheck
citecheck --input draft.txt --report json

It checks each reference against CrossRef, PubMed, Semantic Scholar, and OpenAlex — roughly 240M papers combined. It catches two failure modes human review misses: DOIs that don't resolve to any paper, and papers where the DOI resolves but the claim doesn't match what the paper actually reports. The second type is the harder one. It's what happens when an LLM drafts a sentence from a paper's abstract tone rather than its findings.

CiteCheck also has a GitHub Action version that runs automatically on every manuscript commit if you work in a Git repo. No plugin overhead in your Zotero instance.

I've covered the citation verification workflow in detail in Citation Hallucination: The Silent Killer in AI-Assisted Writing. For where this fits in the broader AI tool stack for clinical researchers, The 2026 AI Tool Landscape for Clinical Researchers maps the full picture.

The Stack That Runs

Better BibTeX (stable export) → Claude Project with .bib as system context → draft manuscript → CiteCheck CLI on export → clean submission copy.

No plugin that promises to read your whole library. No AI that summarizes without being told exactly what to summarize. The Zotero workflow I documented for ARM2 literature screening is in Zotero + Claude: Literature Synthesis Workflow for Systematic Reviews — that post covers the screening layer; this one covers the drafting and verification side.

pip install citecheck — MIT-licensed, 240M-paper index, runs locally. No signup required.