Zotero Agent

English | 中文

Zotero Agent is a Zotero plugin that embeds an MCP (Model Context Protocol) server, turning your local Zotero library into a workspace an AI agent can fully operate — not just read.

It exposes 42 tools spanning library search & retrieval, metadata enrichment, identifier-based import (DOI / arXiv / ISBN / PMID), grey-source PDF download (Sci-Hub / Anna's Archive), duplicate detection & merge, batch tagging, citation-graph expansion, and annotation synthesis — plus an escape-hatch run_javascript for arbitrary in-process automation beyond the built-in tools.

In practice, an AI assistant (Claude, Codex, …) talking to this server can search your library in natural language, clean up metadata and tags in bulk, import and de-duplicate papers, fetch missing PDFs, expand a topic through its citation graph, and synthesize your annotations — with dry-run-by-default safety on every write.

How to Use

Step 1: Install the Zotero Plugin

1. Download the latest zotero-agent.xpi from the Releases Page.
2. Install it in Zotero through Tools -> [gear icon] -> Install Plugin From File....
3. Open Editor -> Settings -> Zotero Agent and enable these permissions:
   1. Allow Remote Access
   2. Enable Write Operations
   3. Run JavaScript (eval)

Note: Run JavaScript (eval) is used to run JavaScript inside Zotero so it can handle tasks that go beyond the built-in MCP tools. Enable it only when you need it.

Step 2: Configure Your AI Client

Get your token from Zotero -> Editor -> Settings -> Zotero Agent -> PSK.

Codex App

Go to Setting -> MCP Servers -> Add server, then use this configuration:

Field	Value
MCP server name	zotero-mcp
Transport	Streamable HTTP
URL	http://127.0.0.1:23120/mcp
Header Key: Authorization	Bearer <YOUR_PSK>
Header Key: Content-Type	application/json

Codex CLI

Edit %USERPROFILE%/.codex/config.toml or ~/.codex/config.toml:

[mcp_servers.zotero-mcp]
enabled = true
url = "http://127.0.0.1:23120/mcp"
http_headers = { Authorization = "Bearer <YOUR_PSK>", "Content-Type" = "application/json" }

Claude Code

$env:ZOTERO_MCP_BEARER_TOKEN = "YOUR TOKEN"

claude mcp add --transport http zotero-mcp http://127.0.0.1:23120/mcp `
  --scope user `
  --header "Authorization: Bearer $env:ZOTERO_MCP_BEARER_TOKEN" `
  --header "Content-Type: application/json"

Note: If you install it inside this project and run it from this project directory, do not use --scope user.

Step 3 (Optional): Recommended Companion Plugins

These plugins run alongside Zotero Agent — they're not bundled in, but once installed, the AI agent can drive them through the run_javascript tool (Zotero's privileged context can reach any installed plugin's API). Install the ones that match your workflow:

Plugin	What it does	How it pairs with this plugin / the agent
jasminum 茉莉花	Scrapes Chinese-database metadata (CNKI / Wanfang / VIP) for Chinese literature with incomplete fields.	This plugin's built-in enrichment targets Western sources (CrossRef / OpenAlex); jasminum fills the Chinese gap. After importing Chinese PDFs, the agent can trigger its scrape / filename-match via run_javascript.
zotero-updateifsE 绿青蛙	Writes impact factor, JCR / CAS quartiles, and other journal metrics into items.	After import_by_identifier / enrich_item_metadata fills the core metadata, the agent can batch-update metrics per collection via run_javascript.
zotero-format-metadata	50+ metadata linters: title case, dates, pages, LTWA journal abbreviations, Chinese name / pinyin cleanup.	Complements enrich_item_metadata (fills fields) with format normalization — the agent can run its lint rules as a post-enrichment cleanup step.
zotero-zotadata	Multi-source metadata fill + multi-provider PDF discovery (Unpaywall / arXiv / CORE / …).	A heavier alternative to find_missing_pdfs when you need broader PDF sources; the agent can invoke its retrieval pipeline via run_javascript.

Let the agent install them for you. With run_javascript (eval) enabled, paste a prompt like this — trim the list to what you need:

Please install these Zotero companion plugins via run_javascript. For each repo: fetch https://api.github.com/repos/<repo>/releases/latest, pick the asset whose name ends in .xpi, install it with AddonManager.getInstallForURL(url) then install.install(), and report each plugin's id / version / active state. Drop any I don't need:

• jasminum (Chinese metadata: CNKI / Wanfang / VIP) — l0o0/jasminum
• Green Frog / updateifsE (impact factor & quartiles) — redleafnew/zotero-updateifsE
• Linter / format-metadata (format cleanup & journal abbreviations) — northword/zotero-format-metadata
• Zotadata (multi-source fill + PDF discovery) — ydeng11/zotero-zotadata

Some plugins may need a Zotero restart to fully activate. Note: most of these are also searchable in the community Zotero Add-on Market plugin (e.g. format-metadata appears there as "Linter for Zotero"); a few (like zotadata) are GitHub-release-only.

See the Built on Open Source — Integration Status & Roadmap section below for how each is called today and which interactions we plan to promote into dedicated MCP tools.

Grey-source PDF Download (Sci-Hub / Anna's Archive)

On top of Zotero's built-in open-access resolvers, the plugin can use Sci-Hub / Anna's Archive as fallback PDF download sources.

Enable in the preferences panel. In Editor -> Settings -> Zotero Agent, turn on the Sci-Hub / Anna's Archive toggle. A source list appears pre-filled with sensible defaults (several Sci-Hub mirrors + Anna's Archive); you can add or remove sources, or restore the defaults. Sources register as manual-only resolvers — used only when you explicitly trigger a download, never in the background. Everything stays off by default.

Download. Once enabled, Zotero's native right-click Find Available PDF automatically includes these sources: Zotero tries free sources first (arXiv / open access) and falls back to the grey sources only when needed. You can also drive it through the MCP tools manage_pdf_resolvers (enable/disable and manage the source list) and find_missing_pdfs (audit which items lack a PDF, then fetch).

Compliance. Sci-Hub / Anna's Archive are grey-area sources. Legal compliance in your jurisdiction is your responsibility.

Development Setup

Clone the repository:

git clone https://github.com/psiQAQ/zotero-agent.git
cd zotero-agent

Set up the plugin development environment:

npm install
npm run build

Load the plugin in Zotero:

# For development with auto-reload
npm run start

# Or install the built .xpi file manually
# The xpi file will be generated at "./.scaffold/build/zotero-agent.xpi"
npm run build

MCP Tools

The table below is based on the actual tool definitions in src/modules/streamableMCPServer.ts.

Tool Name	Purpose
get_libraries	List the Zotero libraries available in the current client.
search_library	Search the Zotero library with filters such as title, year, full text, item type, and relevance scoring.
search_libraries	Search libraries by name.
search_annotations	Search highlights, notes, and comments by query, color, tag, or item scope.
get_item_details	Get detailed metadata for a specific Zotero item.
get_annotations	Get annotations and notes for a specific item or annotation ID.
get_content	Read full-text content from PDFs, attachments, notes, and abstracts.
get_collections	List collections in the library, including recursive tree output when needed.
search_collections	Search collections by name.
get_collection_details	Get detailed information for a specific collection.
get_collection_items	List items inside a specific collection.
get_subcollections	Get child collections under a specific collection.
create_collection	Create a new collection, optionally under a parent collection.
update_collection	Rename or move an existing collection.
delete_collection	Delete a collection, with optional item deletion.
add_items_to_collection	Add one or more items to a collection.
remove_items_from_collection	Remove one or more items from a collection without deleting them from the library.
search_fulltext	Search across cached full-text document content and return matching passages.
get_item_abstract	Get the abstract or summary of a specific item.
semantic_search	Run embedding-based semantic search to find conceptually related content.
find_similar	Find items that are semantically similar to a given item.
semantic_status	Show the status and index statistics of the semantic search service.
fulltext_database	Access the cached full-text database with list, search, get, and stats actions.
write_note	Create, update, or append Zotero notes.
write_tag	Add, remove, or replace tags on Zotero items.
write_metadata	Update item metadata such as title, abstract, DOI, date, or creators.
write_item	Create items, re-parent attachments, or import local files as attachments.
run_javascript	Execute JavaScript inside the Zotero process for advanced automation.
reload_plugin	Reload an installed Zotero plugin for development workflows.
install_plugin_from_url	Install or upgrade a plugin XPI from a reachable URL or file path.
import_by_identifier	Import an item by DOI, arXiv ID, ISBN, or PMID.
find_missing_pdfs	Report items without PDFs or fetch open-access PDFs for them.
manage_pdf_resolvers	Register Sci-Hub / Anna's Archive into Zotero's native PDF resolver pref (grey sources default automatic=false, manual-only); actual download is via find_missing_pdfs.
extract_identifier_from_pdf	Mine DOI or arXiv ID from a PDF's fulltext cache using frequency voting. Read-only.
find_doi	Reverse-lookup a DOI via CrossRef title-similarity (≥0.86 threshold); dry-run default, confirm write requires write.enabled.
enrich_item_metadata	Fill missing fields (abstract/venue/volume/issue/pages/ISSN/publisher/date) from a DOI via doi.org CSL-JSON + OpenAlex; dry-run default, confirm write requires write.enabled.
check_retractions	Check items against scite.ai editorial notices such as retractions or corrections.
find_related_papers	Traverse the citation graph through OpenAlex to find citing or referenced papers.
synthesize_annotations	Aggregate highlights and notes into a literature-review-oriented markdown bundle.
find_duplicates	Detect duplicate items using Zotero's native duplicates engine.
merge_duplicates	Merge duplicate items into a chosen master item.
batch_update_tags	Run bulk tag operations such as add, remove, or rename.

Built on Open Source — Integration Status & Roadmap

This plugin stands on the shoulders of several open-source projects: it forks a writable in-process MCP base and folds in the best ideas from ~17 reference projects (archived read-only as submodules under refs/). Full technical comparisons live in:

• AI / MCP integration approaches — 5 projects
• Metadata enrichment plugins — 8 projects
• PDF download approaches — 4 projects

1. What this fork adds on top of the base

Forked from cookjohn/zotero-mcp — a clean, in-process, locally-writable MCP server (27 tools, direct saveTx/eraseTx). Our additions:

Area	cookjohn base	This fork adds
Auth	loopback only	PSK Bearer auth + Origin validation (DNS-rebind defense)
Eval	none	run_javascript privileged eval tool (timeout + 100 KB cap)
Tools	27	42 — +15 tools (identifier import, missing-PDF audit, citation graph, dedup, batch tags, metadata enrich, DOI reverse-lookup, grey-source download, …)
Search	keyword	+ RRF hybrid semantic search, 0-result fallback ladder
Testing	none	26-scenario in-process selfTest + node unit tests
Deploy	manual	one-shot deploy-live (base64 ship + self-upgrade) + reload_plugin / install_plugin_from_url
i18n	zh / en	+ de / es / fr / ja
CJK	—	byte-level HTTP read fix (dense CJK bodies no longer mojibake)

2. Capabilities absorbed from reference projects (integrated)

Ideas were re-implemented (not vendored) into native tools:

AI / MCP — see comparison

Source	Absorbed into	Status	TODO
54yyyu/zotero-mcp	import_by_identifier, find_missing_pdfs, find_related_papers (OpenAlex), check_retractions (scite), synthesize_annotations	✅	port more of its 62-tool breadth (BibTeX / CSL import, batch OA indexing)
introfini/ZotSeek	RRF hybrid semantic_search	✅	WebGPU acceleration; Matryoshka dim truncation
introfini/mcp-server-zotero-dev	run_javascript, reload_plugin, install_plugin_from_url	✅	screenshot / DOM-inspection tools for UI debugging

Metadata — see comparison

Source	Absorbed into	Status	TODO
zotero-metadata-hunter	enrich_item_metadata (field-level fill from DOI CSL-JSON + OpenAlex)	✅	preprint → published-version upgrade
zotero-doi-fix	find_doi (title-similarity fusion), extract_identifier_from_pdf	✅	DOI repair (bad → revalidate → replace w/ backup), not just reverse-lookup

PDF — see comparison

Source	Absorbed into	Status	TODO
pdferret, zotero-scipdf	manage_pdf_resolvers (native findPDFs.resolvers read/write) + grey-source download	✅	multi-mirror rotation; stronger DOI extraction (scipdf's 5-regex + attachment scrape)

3. Not integrated, but callable via run_javascript (interaction boundary + roadmap)

These plugins aren't folded into this one, but if the user installs them, an AI agent can drive them through run_javascript — which runs in Zotero's privileged context and can reach any installed plugin's exposed API. Current boundary and future dedicated-tool directions:

Plugin	Capability	Current agent boundary (via run_javascript)	TODO (dedicated tool)
jasminum 茉莉花	Chinese metadata scraping (CNKI / Wanfang / VIP)	Invoke its scrape / filename-match functions on the Zotero.Jasminum namespace if exposed	wrap fetch_chinese_metadata(itemKey) — one call to scrape + fill CNKI fields
zotero-updateifsE 绿青蛙	Impact factor / JCR & CAS quartiles	Call its easyScholar update path per item	update_journal_metrics(scope) tool
zotero-format-metadata	50+ format linters, LTWA journal abbreviation	Trigger its lint rules programmatically	lint_metadata(scope, rules) tool
zotero-zotadata	Multi-source fill + multi-provider PDF discovery	Call its retrieval pipeline	deep_enrich(itemKey) combining fill + PDF fetch

run_javascript is the universal escape hatch: any installed plugin exposing functions on the Zotero object (or a global) can be driven by the agent today. The TODO column is about promoting the most-used interactions into typed, dry-run-safe MCP tools.