zotero-library-mcp

An MCP server that lets Claude, Codex, and ChatGPT add papers and books to your Zotero library by DOI, arXiv ID, or ISBN — and manage your collections, tags, and items.

The server supports both MCP transports used by these clients:

• stdio (default) for Claude Code, Claude Desktop, Codex CLI/IDE, and the ChatGPT desktop app

• Streamable HTTP for a hosted ChatGPT app or any remote MCP client

Tools

Adding papers

• add_paper_by_doi — Resolve a DOI via CrossRef and add the paper to Zotero (with duplicate detection)

• add_papers_by_dois — Batch-add up to 50 papers at once

• add_paper_by_arxiv_id — Add a preprint by arXiv ID (uses DOI when available, falls back to arXiv metadata)

• add_item_from_metadata — Create any supported Zotero item type from validated manual metadata

Adding books

• add_book_by_isbn — Resolve an ISBN via Open Library and add the book to Zotero (with duplicate detection)

Searching & browsing

• list_libraries — Discover the API key owner's personal library and shared group libraries, with IDs and key permissions

• search_library — Search your Zotero library by title, author, tag, etc., paginated via start/limit (falls back to fuzzy matching when the exact search returns no results)

• get_item_details — View full metadata for any item

• get_recent_items — List recently added items

• get_unfiled_items — Get items not in any collection

• search_fulltext — Search Zotero metadata and indexed full text

• find_duplicates — Find duplicate items by DOI, ISBN, or normalized title

• list_attachments — List every attachment and choose a specific PDF key

• health_check — Verify library credentials, access, and storage configuration

Reading & annotating

• get_item_fulltext — Return bounded plain text from Zotero's index or a PDF, without leaking temporary paths

• get_bibtex — Read-only BibTeX/BibLaTeX export for items, a collection, or the full library

• save_bibtex — Save an export to an authorized local path

• get_annotations — List all highlights and annotations on a paper's PDF

• create_annotation — Highlight a text passage in a PDF (searches for the exact text, creates a visible highlight in Zotero's reader, and returns a preview image for verification). Smart overlap handling: exact duplicates update the existing comment; sub-passages get a contrasting highlight color automatically.

• add_note — Add a note to an item

• list_notes, update_note, delete_note — Manage existing notes

• update_annotation, delete_annotation — Edit or remove annotations

File attachments

• attach_file — Attach a local file over stdio or a ChatGPT file input over HTTP

• download_pdf — Return a remote-safe MCP file resource

• save_pdf — Save a PDF to an authorized local path

Collections

• list_collections — List all collections (with nesting)

• create_collection — Create a new collection (optionally nested under a parent)

• get_collection_items — Browse items in a collection, paginated via start/limit

• add_to_collection — Add an existing item to a collection

• remove_from_collection — Remove an item from a collection (keeps it in your library)

• rename_collection, move_collection — Reorganize collections

Tags

• list_tags — List all tags in your library

• add_tags — Add one or more tags to an item (with optional color)

• remove_tags — Remove tags from an item

• delete_tags — Delete tags from the entire library

• set_tag_color — Assign a color to a tag (appears in Zotero's tag selector)

• rename_tag — Rename a tag across all items in your library

• unset_tag_color — Remove a tag color without deleting the tag

Verification

• verify_items — Re-check recent items against CrossRef to catch bad DOIs or title mismatches

Deleting

• delete_item — Permanently delete an item from your library

• delete_collection — Permanently delete a collection

• trash_item, restore_item — Prefer reversible trash operations for ordinary cleanup

The server also exposes the standard read-only search and fetch tool shapes used by ChatGPT company knowledge and deep research.

Related MCP server: zotero-mcp-lite

Prerequisites

1. A Zotero account

2. A Zotero API key with write permissions: https://www.zotero.org/settings/keys

3. Your Zotero library ID (shown on the same page, or in your profile URL)

4. uv installed

Quick Start

Codex and the ChatGPT desktop app

Codex and the ChatGPT desktop app share MCP configuration on the same Codex host. Add the server once:

codex mcp add zotero \
  --env ZOTERO_LIBRARY_ID=your_library_id \
  --env ZOTERO_API_KEY=your_api_key \
  -- uvx --from git+https://github.com/RaulSimpetru/zotero-library-mcp zotero-mcp

Then restart Codex or the ChatGPT desktop app. In Codex, use /mcp to confirm that zotero is connected. In ChatGPT desktop, open Settings → MCP servers to view the same server.

For WebDAV storage, add the three ZOTERO_WEBDAV_* values shown in the WebDAV example. If the desktop app cannot find uvx, replace it with the full path returned by which uvx.

You can also configure the server directly in ~/.codex/config.toml:

[mcp_servers.zotero]
command = "/full/path/to/uvx"
args = ["--from", "git+https://github.com/RaulSimpetru/zotero-library-mcp", "zotero-mcp"]
env_vars = ["ZOTERO_LIBRARY_ID", "ZOTERO_API_KEY", "ZOTERO_LIBRARY_TYPE", "CROSSREF_MAILTO", "ZOTERO_WEBDAV_URL", "ZOTERO_WEBDAV_USER", "ZOTERO_WEBDAV_PASSWORD"]
startup_timeout_sec = 30
tool_timeout_sec = 120

With env_vars, start Codex/ChatGPT from an environment that contains those variables. Use [mcp_servers.zotero.env] instead if you intentionally want to store their values in the config file.

Claude Code

Install it as a plugin (recommended):

/plugin install zotero --marketplace RaulSimpetru/zotero-library-mcp

On older Claude Code versions, run /plugin marketplace add RaulSimpetru/zotero-library-mcp first, then /plugin install zotero@zotero-library-mcp. Claude Code asks for your library ID and API key (plus optional WebDAV and Unpaywall settings) when you enable the plugin, and keeps the API key and WebDAV password in your system keychain. To get a new release, run /plugin marketplace update zotero-library-mcp, or turn on auto-update for the marketplace in /plugin.

Plugin tools are named mcp__plugin_zotero_zotero__<tool> rather than mcp__zotero__<tool>; update any permission rules that name the old tools. If you previously added the server with claude mcp add, remove that entry (claude mcp remove zotero -s user) so the two copies don't both run.

Or add it as a plain MCP server:

claude mcp add zotero \
  -e ZOTERO_LIBRARY_ID=your_library_id \
  -e ZOTERO_API_KEY=your_api_key \
  -- uvx --from git+https://github.com/RaulSimpetru/zotero-library-mcp zotero-mcp

WebDAV setup

To use WebDAV file storage (e.g. Synology, Nextcloud), include the WebDAV variables:

claude mcp add zotero \
  -e ZOTERO_LIBRARY_ID=your_library_id \
  -e ZOTERO_API_KEY=your_api_key \
  -e ZOTERO_WEBDAV_URL=https://your-webdav-server.com \
  -e ZOTERO_WEBDAV_USER=your_username \
  -e ZOTERO_WEBDAV_PASSWORD=your_password \
  -- uvx --from git+https://github.com/RaulSimpetru/zotero-library-mcp zotero-mcp

Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "zotero": {
      "command": "/full/path/to/uvx",
      "args": ["--from", "git+https://github.com/RaulSimpetru/zotero-library-mcp", "zotero-mcp"],
      "env": {
        "ZOTERO_LIBRARY_ID": "your_library_id",
        "ZOTERO_API_KEY": "your_api_key",
        "ZOTERO_WEBDAV_URL": "https://your-webdav-server.com",
        "ZOTERO_WEBDAV_USER": "your_username",
        "ZOTERO_WEBDAV_PASSWORD": "your_password"
      }
    }
  }
}

Note: Claude Desktop doesn't inherit your shell's PATH, so you need the full path to uvx. Find it with which uvx in your terminal.

ChatGPT on the web (Apps SDK / developer mode)

ChatGPT web connects to an HTTPS Streamable HTTP endpoint. Start the server locally with the HTTP transport, then make it reachable through Secure MCP Tunnel or another authenticated HTTPS deployment:

ZOTERO_LIBRARY_ID=your_id ZOTERO_API_KEY=your_key \
  uvx --from git+https://github.com/RaulSimpetru/zotero-library-mcp zotero-mcp \
  --transport streamable-http \
  --host 127.0.0.1 \
  --port 8000 \
  --allowed-host your-tunnel.example.com

The MCP endpoint is https://your-tunnel.example.com/mcp. Enable developer mode in ChatGPT, create a developer-mode app, and enter that URL as the MCP server URL. See OpenAI's Connect from ChatGPT guide for the current UI flow.

Security: The safest personal setup is OpenAI Secure MCP Tunnel with the MCP server bound to loopback. HTTP mode disables all server-path reads and writes by default. attach_file accepts ChatGPT's authorized file object, while download_pdf returns an opaque MCP resource link. Safety annotations are approval hints, not an authorization boundary.

For a public deployment, configure an external OAuth 2.1 identity provider. The server validates JWT access tokens against its JWKS endpoint:

export ZOTERO_MCP_OAUTH_ISSUER=https://auth.example.com
export ZOTERO_MCP_OAUTH_RESOURCE=https://zotero.example.com
export ZOTERO_MCP_OAUTH_JWKS_URL=https://auth.example.com/.well-known/jwks.json
export ZOTERO_MCP_OAUTH_SCOPES=zotero:read,zotero:write

The authorization server must publish OAuth/OIDC discovery metadata, support the MCP OAuth 2.1 flow with PKCE, issue tokens for ZOTERO_MCP_OAUTH_RESOURCE, and include the configured scopes. See OpenAI's authentication guide. For testing behind an already authenticated gateway only, --allow-unauthenticated-http explicitly acknowledges an unauthenticated non-loopback listener.

This process still uses one server-side Zotero API key. Every authenticated MCP client can target libraries available to that key, including shared groups. Use a key scoped to the libraries intended for those clients. A true multi-user service must map the verified OAuth identity to separate Zotero credentials and enforce per-user authorization; that deployment architecture is intentionally outside this personal-server package.

If an HTTP deployment genuinely needs server paths, enable them only inside confined roots:

zotero-mcp --transport streamable-http \
  --allow-server-files \
  --file-root /srv/zotero-mcp/exports

HTTP launch settings can also be supplied as environment variables:

Run standalone

ZOTERO_LIBRARY_ID=your_id ZOTERO_API_KEY=your_key \
  uvx --from git+https://github.com/RaulSimpetru/zotero-library-mcp zotero-mcp

Personal and shared group libraries

The environment variables select the default library. Existing calls that omit library arguments continue to use that default.

Call list_libraries() to discover the key owner's personal library and shared groups. Results include name, library_id, library_type, is_default, and key_permissions. Group results are paginated with limit and start; follow next_start until it is null. Discovery also works when the default is a group.

Every other tool accepts optional library_id and library_type arguments:

list_libraries()
list_collections(library_id="123456", library_type="group")
search_library(query="ultrasound", library_id="123456", library_type="group")
get_item_details(item_key="<key from that group>", library_id="123456", library_type="group")
search_library(query="ultrasound")  # Uses the configured default again

Pass both arguments to select a library, or omit both. Use IDs returned by list_libraries, rather than group names. Selection applies to one call and never changes the environment or another client's target, including concurrent calls. Item and collection keys must come from the selected library. Zotero enforces the API key's permissions and the user's group rights; a denied group request fails without falling back to the personal library. health_check accepts the same arguments to verify a specific library without writing to it.

Group attachments use Zotero's built-in storage. WebDAV is used only for the configured personal library, even when the same server accesses shared groups.

Environment Variables

Note: If all three ZOTERO_WEBDAV_* variables are set, attachments in the configured personal library use WebDAV instead of Zotero's built-in storage. Group attachments always use Zotero storage. The server automatically appends /zotero to the WebDAV base URL, matching Zotero Desktop's behavior.

Version 0.10.2

• Install as a Claude Code plugin: /plugin install zotero --marketplace RaulSimpetru/zotero-library-mcp.

• An empty ZOTERO_LIBRARY_TYPE now means the default (user), as plugin installs pass unset optional settings as empty strings.

Version 0.10.1

• Attaching files to Zotero storage works again (group libraries and setups without WebDAV): Zotero rejects the full local path Pyzotero sent as the filename. Uploads are now confirmed, and a rate-limited or failed upload is reported instead of leaving an empty attachment.

• list_tags returns every tag; Zotero's tag paging metadata is wrong past 200 tags, which made the old listing drop the rest.

• Leftover query parameters no longer leak between calls (e.g. a duplicate check after fetching an item template only searched items of that type).

• Searches and tag renames containing the word "tags" no longer crash.

• Tagging, filing, and editing notes or annotations works on trashed items.

• Highlights land correctly on rotated pages with offset crop boxes, match whole words, check for duplicates at the same location only, and report a missing occurrence instead of highlighting a different passage.

• arXiv imports detect existing copies of any version of the paper.

• CrossRef imports keep organization authors and editors and map books, edited books, chapters, and dissertations to the right item types; Open Library books get their authors from the work record.

• verify_items handles DataCite DOIs (arXiv, Zenodo), HTML in CrossRef titles, and subtitles; health_check reports read-only keys as write_access: false.

• BibTeX exports keep Zotero's order and unique citekeys across pages.

• Lists of child items (attachments, notes, annotations) are no longer capped at 100; linked PDFs are skipped, and standalone PDFs can be read.

• Removing the last tag color works; deleting tags containing || is refused.

• Hardening for hosted HTTP deployments: downloads connect to the validated IP (DNS rebinding), JWKS refetches are rate-limited, and file saves use exclusive temporary files.

• DOI imports no longer create duplicates. Zotero's search does not index DOIs, so the old check never matched; imports now search the resolved title and compare DOIs (including DOI: lines in Extra).

• ISBN imports no longer create duplicates of books stored with hyphenated ISBNs; ISBN-10 and ISBN-13 forms now match each other.

• search_library returns papers instead of matching attachments and notes.

• Rate-limited writes are never retried (a retry could duplicate the item) and now fail instead of reporting success.

• add_papers_by_dois(attach_pdfs=True) downloads PDFs in parallel.

• search_fulltext and search query top-level items (about 20x faster).

• Fuzzy search and find_duplicates reuse a library cache that is kept current with Zotero's incremental sync (since=), so repeat calls take well under a second.

• CrossRef requests use UNPAYWALL_EMAIL as the polite-pool contact when CROSSREF_MAILTO is unset, and retry rate-limited requests.

• create_annotation highlights quotes that wrap across lines in full, counts occurrence across pages, and uses the PDF's own page labels.

• PDFs attached during DOI/arXiv import are named after the item title, and the result now says why a PDF was not attached.

• List output shows the correct year for dates like 12/1/1993.

• health_check returns library_type as user/group and reports whether open-access PDF lookup (UNPAYWALL_EMAIL) is configured.

• New list_trash tool for finding keys to pass to restore_item.

• Faster get_bibtex for item keys, rename_tag, get_recent_items, fuzzy searches, and find_duplicates; Zotero API connections are reused across calls.

• Full-library and collection get_bibtex exports stop downloading once max_chars is exceeded.

• get_item_fulltext prefers a PDF's index over a web snapshot's.

• find_duplicates reports items_scanned and scan_capped.

• search falls back to the item's zotero.org link when it has no URL.

• Shorter tool descriptions; library targeting is explained once in the server instructions.

Version 0.10.0

• Discover personal and shared libraries with list_libraries.

• Target a library on each tool call using library_id and library_type; existing calls keep the configured default.

• Keep concurrent calls and nested exports scoped to the selected library.

• Use Zotero storage for shared-library attachments and personal WebDAV only for the configured personal library.

• Keep current dependency imports and HTTP logging compatible with MCP stdio.

Restart the MCP connection after upgrading to load the new tool schemas.

Upgrading to 0.8

Two path-writing operations were split from their read-only counterparts so remote clients can apply correct safety approvals:

• get_bibtex(save_path=...) is now save_bibtex(save_path=...); get_bibtex only returns data.

• download_pdf(save_path=...) is now save_pdf(save_path=...); download_pdf returns an opaque MCP resource link.

Existing read-only calls to get_bibtex and download_pdf continue to work.

0.8.1

• PDF and annotation-preview resources now keep their precise MIME type when read by clients.

• PDF downloads report progress and return a clear WebDAV timeout error.

• health_check reports API-key write permission without modifying the library.

• Server instructions are shorter, reducing repeated client context usage.

How it works

1. You provide a DOI, arXiv ID, or ISBN

2. The server queries the appropriate API to get full metadata:

   • DOI → CrossRef API

   • arXiv ID → arXiv API (with CrossRef fallback when a DOI exists)

   • ISBN → Open Library API

3. Metadata is mapped to Zotero's item format (title, authors, journal/publisher, date, etc.)

4. The item is created in your Zotero library via the Zotero Web API

License

MIT

mcp-name: io.github.RaulSimpetru/zotero-library-mcp