Zotero MCP with Claude and Codex: A Paradigm Shift for Academic Writing Workflows

Forum topic · 小凯 · 2026-06-07

Summary

Large language models often fabricate plausible-looking citations, a fatal flaw for academic writing. Zotero MCP (Model Context Protocol) addresses this by enforcing a shift from generative citations to retrieval-based citations: the AI can only reference literature that the user has personally verified and stored in their Zotero library. The article explains how Zotero 7 can be configured to expose metadata, PDFs, and notes to AI clients such as Claude Desktop and Codex via MCP, then walks through a three-step workflow involving folder structure, MCP installation, and Academic Research Skills (ARS) prompts. A key insight is that converting papers to Markdown for Obsidian yields deeper content understanding than relying on raw metadata. Limitations such as library coverage, citation format conversion, and AI over-citation are discussed, alongside a recommended toolchain combining Zotero Connector, Better BibTeX, Zotero MDNotes, and Obsidian for a verifiable, traceable, and reusable academic writing infrastructure.

Key Points

Citation hallucination is an integrity issue, not a style issue. LLMs like Claude 3.5 routinely fabricate author names, years, or misattribute findings, which is especially dangerous for non-native writers who may not verify originals.
MCP shifts AI from passive file consumption to active querying. The Model Context Protocol (Anthropic, 2024) lets AI assistants query external tools in real time, enabling retrieval-based rather than generative citations.
Zotero MCP capabilities and limits. It can search by keyword, author, tag, or folder; read metadata (title, authors, journal, DOI, abstract); access PDFs when available; insert formatted citations; and surface highlights and notes. By default only metadata is exposed, and quality depends entirely on what the user has ingested.
Three-step workflow.

Zotero-side setup:

AI-side MCP install:

zotero

{"command": "npx", "args": ["-y", "zotero"]}

Combine with Academic Research Skills (ARS):

Markdown-to-Obsidian is superior to direct Zotero access. Converting papers via tools like zotero-mdnotes gives the AI full text, personal annotations, highlights, and semantic search, versus shallow metadata matching.
Known limitations. Coverage is bounded by the user's library; citation styles (APA, MLA, IEEE, GB/T 7714) require CSL configuration; AI tends to over-cite unless constrained by ARS prompts; and version drift between drafts and the library requires locking the library state and re-verifying before submission.
Recommended toolchain layers.
Acquisition: Zotero Connector, Zotero PDF Translate
Management: Zotero 7, Better BibTeX, Zotero Style
Content extraction: Zotero MDNotes, Obsidian Zotero Integration
Writing: Claude Desktop / Codex + Zotero MCP, ARS prompts
Verification: Zotero's built-in bibliography check plus human review
Future directions. Unified retrieval across Zotero, arXiv, Semantic Scholar, and PubMed; automatic library syncing; collaborative shared libraries; and post-hoc AI audits flagging cited-but-undiscussed sources.

References & Tools

Zotero MCP: https://github.com/topics/zotero-mcp
PapersFlow MCP: https://doxa.papersflow.ai/mcp
Zotero MDNotes: https://github.com/argenos/zotero-mdnotes
Obsidian Zotero Integration: https://github.com/mgmeyers/obsidian-zotero-integration
Academic Research Skills (ARS): referenced from prior coverage

Bottom Line

The core value of Zotero MCP is not making AI read faster but making AI cite honestly. By binding generation to a user-verified corpus, the workflow upgrades academic integrity infrastructure rather than raw throughput, which matters most for non-native writers, graduate students, and anyone producing literature-heavy work.