
Zotero Native MCP
io.github.dvdsosaPublished by GitHub account dvdsosav1.0.2Updated Oct 26, 2026
Read and write your local Zotero 10 library. No plugin, no zotero.org API key, no cloud.
The publisher's own server: NativeExpress
Overview
Local MCP server that lets an assistant search, read, create, edit and delete items in a Zotero 10 library, including PDF full text, entirely offline.
- What you can do
- Search and read items, item children, tags, saved searches and PDF full text
- Create, update, delete and restore items, and add or remove them from collections
- Create, update, delete and restore collections
- Attach local PDF files and look up attachment paths
- List and empty the trash, export items, and check status, libraries and item type fields
- When to use it
- Useful when you keep references in a local Zotero 10 library and want an assistant to find sources, pull PDF text, file papers into collections or add references and attachments without a zotero.org account or plugin.
- Requirements
- Zotero 10 or newer running locally, with "Allow other applications on this computer to communicate with Zotero" enabled in Zotero's Advanced settings. Node.js 22 or newer for the npm package. Optional environment variables ZOTERO_LOCAL_APP_NAME and ZOTERO_LOCAL_PORT (default 23119). No API key or account.
Installation
In SourceWeft
- Open Zotero Native MCP in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
zotero-native-mcp
An MCP server that reads and writes your Zotero 10+ library, entirely offline.
[Zotero 10+] [npm] [CI] [License: MIT] [DOI]
Connect Zotero to Claude, Claude Code, Cursor, or any Model Context Protocol client. Your assistant can search your library, read PDF full text, create collections, add references, and attach PDFs from your disk.
Every operation runs against the Zotero application on 127.0.0.1. No
zotero.org account. No web API key. No Zotero plugin. No cloud round trip.
Why this exists
Zotero 10 added write support to its built-in local API. Before that, every
Zotero MCP server had to work around a read-only local endpoint, either by
routing writes through api.zotero.org (slow, needs an API key, needs your
library synced to the cloud) or by shipping a separate Zotero plugin you had to
install and keep up to date.
This server uses the native capability directly. Nothing to install inside Zotero, no credentials to manage, and reads land in 8–60 ms because nothing touches the network.
Requirements
- Zotero 10 or newer, running. This is a hard floor, not a preference: writing through the local API did not exist before Zotero 10. On Zotero 7–9 the read tools work and every write fails.
- Zotero → Settings → Advanced → enable "Allow other applications on this computer to communicate with Zotero".
- Node.js 22 or newer. Node 20 reached end of life in April 2026.
Tested on
Everything below is a statement of evidence, not of intent. Other platforms
and clients are expected to work. The server is portable TypeScript talking to
127.0.0.1, with nothing platform-specific by design. But they have not been
verified, and this table is the honest extent of it.
CI runs the unit and mock-protocol suites across a matrix of Linux, macOS and
Windows × Node 22, 24 and 26, so portability of the code itself is covered on
all three platforms. What no runner can cover is the conversation with a real
Zotero, since none is installed there. That part was done by hand:
scripts/coverage.mjs reaches all 28 tools on all three operating systems,
against a live library.
Not yet verified. Intel Macs; Claude Code on Linux, where only the scripts were run; Claude Desktop, Cursor and other MCP clients anywhere.
If you run it somewhere not on this list, a report either way is welcome, those are the most useful issues this project can receive right now.
Back up your library first
[!WARNING] This server can modify and delete items in your Zotero library. Back it up before you start, and keep backing it up.
Deleting is reversible by default:
zotero_delete_itemsandzotero_delete_collectionmove things to Zotero's trash, where you can restore them from the Zotero window or withzotero_restore_items. But Zotero empties that trash automatically after 30 days, and both tools take apermanent: truethat erases outright, no undo, attachment files removed from disk, nothing in the API able to bring them back.zotero_empty_trashdoes the same to everything already in the trash.These tools are driven by an assistant interpreting instructions in natural language, which can misread which item you meant.
To back up: quit Zotero, then copy your whole data directory,
~/Zoteroon macOS and Linux,%USERPROFILE%\Zoteroon Windows, or whatever Settings → Advanced → Files and Folders reports. It holdszotero.sqliteand thestoragefolder with every attachment. Zotero's own guidance is at zotero.org/support/zotero_data.Syncing to zotero.org is not a backup: a deletion syncs too.
This software is provided as is, without warranty of any kind, and its authors accept no liability for data loss. See LICENSE.
Quick start
Other clients (Claude Desktop, Cursor, …)
--scope user registers it for every project. Without it, claude mcp add
defaults to the local scope, which ties the server to the one directory you
ran the command in, open anything else and the tools are simply absent.
No environment variables are needed. Start a new session, then ask your
assistant to run zotero_status to confirm the connection.
The first time a tool writes, Zotero shows a dialog asking whether to allow it. Choose "Always Allow" so you are not asked again.
Documentation
Tools at a glance
Collections: list_collections get_collection create_collection
update_collection delete_collection restore_collection
Items: search_items get_item get_item_children create_items
update_item delete_items restore_items add_items_to_collection
remove_items_from_collection get_item_fulltext export_items
Trash: list_trash empty_trash (and the restore_* tools above)
Attachments: attach_file get_attachment_path
Discovery: list_tags list_saved_searches run_saved_search
System: status authorize list_libraries get_item_type_fields
All names are prefixed zotero_. See the reference for
full signatures.
Prior art
This project is not a fork. It was written from scratch once Zotero 10 made native local writes possible, but it stands on the shoulders of earlier work that solved the same problem under tighter constraints:
- 54yyyu/zotero-mcp: the most
widely used Zotero MCP server. Rich feature set including semantic search;
writes go through
api.zotero.org. - cookjohn/zotero-mcp: a Zotero 7 plugin exposing an MCP endpoint from inside Zotero, with vector search.
- Ayanya-0628/zotero-mcp and
dzackgarza/zotero-local-write-api: local-first writes via a companion
.xpiwrite endpoint. - kujenga/zotero-mcp: a lightweight Python server for the Zotero API.
If you need Zotero 7/8/9 support, semantic or vector search, or writes to a library you only have cloud access to, then those projects remain the right choice.
Contributing
Issues and pull requests are welcome. CI must pass on Linux, macOS and Windows across Node 22, 24 and 26.
See CONTRIBUTING for the development setup, the test suite, and the scripts that exercise every tool against a live Zotero.
License
MIT © David Sosa
Keywords: Zotero MCP server · Model Context Protocol · Zotero Claude integration · Zotero local API · offline reference manager automation · Zotero AI assistant · BibTeX export · academic research tooling · Claude Code Zotero · Cursor Zotero
Source: README.md at commit 3a612c8
Tools
0Version history
2- v1.0.2LatestOct 7, 2026
- v1.0.1Sep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts
Other Native integrations
NativeExpress
express.native.wwwOfficial publisher
NativeExpress's hosted MCP server, published by www.native.express; no sign-in required; Streamable HTTP; 5 tools.
Purple8 — AI-Native Application Backend
ai.purple8.www
Remote MCP backend that gives an AI assistant 176 tools to build and run app data, auth, workflows, SLAs, human-in-the-loop and RAG.

Ruby Native
Published by GitHub account ruby-native
Local MCP server for the Ruby Native Rails gem that checks native view signals and config and drives app builds, listings and store submission.

Pages Native
Published by GitHub account viames
Native Apple Pages: guarded edits, templates, batches, recipe history and PDF previews.

Numbers Native
Published by GitHub account viames
Lets an assistant create, inspect and edit real Apple Numbers spreadsheets on macOS through the Numbers app, with formulas, charts and CSV.