Zotero Fulltext

MCP Server PyPI License: MIT

Access your Zotero library with your favorite AI tool.

This MCP server for Zotero 8+ gives Claude and Codex access to your library via search and citekeys. It talks directly to Zotero's local API and aims to keep token usage low. Fulltext is fetched only on demand.

Quick Start

1. Make sure Zotero 8+ is running with the local API enabled (Settings → Advanced → "Allow other applications on this computer to communicate with Zotero").
2. Install the Claude Code plugin:

/plugin marketplace add statzhero/zotero-fulltext
/plugin install zotero@zotero-fulltext

3. Run /mcp to confirm the server is connected, then try a slash command:

/zotero:find sustainability reporting

Also works with Claude Desktop, Codex, and as a standalone MCP server without slash commands.

Commands

The Claude Code plugin provides four slash commands:

Command	What it does
/zotero:find <query>	Search the whole library
/zotero:lookup <citekey>	Exact citekey metadata lookup
/zotero:read <citekey>	Numbered fulltext paragraphs
/zotero:within <citekey> <query>	Search inside one paper's fulltext

find searches the whole library. lookup is lightweight metadata confirmation. read returns the actual paper text. within searches only one item's indexed fulltext.

Without the plugin, all the same functionality is available through the MCP tools directly (see Tools).

Installation

Claude Code (plugin with slash commands)

/plugin marketplace add statzhero/zotero-fulltext
/plugin install zotero@zotero-fulltext

This installs the MCP server and slash commands. Run /mcp to confirm the server is connected.

Claude Desktop

Requires uv (install with brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh).

Open Claude Desktop → Settings → Developer → Edit Config, and add:

{
  "mcpServers": {
    "zotero": {
      "type": "stdio",
      "command": "uvx",
      "args": ["zotero-fulltext"]
    }
  }
}

Save the file, then fully quit and reopen Claude Desktop (closing the window is not enough). Open a new chat and confirm the server is available.

Codex (plugin with skills)

Note: The Codex plugin marketplace is still rolling out. The install flow below may change.

codex plugin marketplace add statzhero/zotero-fulltext
codex plugin install zotero

This installs the MCP server and four skills (find, lookup, read, within).

Codex (manual MCP only)

Requires uv (install with brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh).

Add to ~/.codex/config.toml:

[mcp_servers.zotero]
command = "uvx"
args = ["zotero-fulltext"]

Restart Codex, then run codex mcp list to verify the server appears.

Design

The server is intentionally simple and read-only. It relies on Zotero's own search index rather than building a second one.

• Startup builds a metadata index mapping citekeys to items and attachments.
• Library changes are tracked with Zotero version headers and incremental sync.
• Fulltext is fetched only on demand and cached in memory (TTL/LRU).
• All outputs are bounded by default: 10 search hits, 80 paragraphs with a character budget, 20 fulltext matches.
• Item results include item_uri and fulltext_uri so clients can attach standard zotero://... resources directly.
• Creator roles (author, editor, translator, etc.) are preserved and grouped in results.
• If a lookup finds no citekey, it returns found=false. If a search finds nothing, it returns results=[]. There is no web fallback.

Tools

The server exposes five MCP tools. The slash commands above are convenience wrappers.

lookup(citekey)

Exact citekey lookup. Citekeys are resolved in order:

1. Native Zotero 8 citationKey
2. Legacy Better BibTeX Citation Key: line in Extra
3. Deterministic generated fallback

If an item later gains a real citekey, the generated key is kept as an alias.

search(query, collection?, tag?, limit?)

Searches Zotero with qmode=everything, collapses attachment hits to parent items, and ranks exact citekey matches first.

collections()

Lists collections in the current library.

fulltext(citekey, offset?, limit?)

Fetches indexed attachment fulltext, splits it into numbered paragraphs, and returns a bounded slice (default: 80 paragraphs). Large extracted paragraphs are split into smaller chunks, and each response has a soft character budget. If the response includes truncated=true, request the same citekey again with offset=next_offset to continue reading.

Fulltext responses include paging metadata:

Field	Meaning
paragraph_count	Total available paragraph chunks for the item
returned_count	Number of paragraph chunks in this response
returned_chars	Approximate text characters returned
max_chars	Character budget used for this response
truncated	Whether more paragraph chunks remain
next_offset	Offset to pass into the next fulltext call, or null when complete

fulltext_search(citekey, query, before?, after?, limit?)

Searches within a single item's paragraphized fulltext and returns matching paragraphs with surrounding context.

Environment Variables

By default the server connects to a local personal library with no authentication. Set these variables to change that:

Variable	Default	Description
ZOTERO_LIBRARY_TYPE	user	user for personal libraries, group for group libraries
ZOTERO_LIBRARY_ID	0	Zotero user or group ID (required for group libraries)
ZOTERO_API_KEY	—	API key for authenticated or remote access
ZOTERO_API_BASE_URL	http://127.0.0.1:23119/api	Base URL for the Zotero API
ZOTERO_MAX_PARAGRAPH_CHARS	1800	Maximum characters per returned fulltext chunk
ZOTERO_MAX_FULLTEXT_CHARS	60000	Soft character budget for each fulltext response; continue with next_offset

Example for a group library in Claude Code:

{
  "mcpServers": {
    "zotero": {
      "type": "stdio",
      "command": "zotero-fulltext",
      "env": {
        "ZOTERO_LIBRARY_TYPE": "group",
        "ZOTERO_LIBRARY_ID": "12345"
      }
    }
  }
}

Related Projects

Other MCP servers for Zotero, with different design goals:

• 54yyyu/zotero-mcp — Feature-rich: read-write operations, optional semantic search via ChromaDB, Web API support. Heavier dependencies.
• kujenga/zotero-mcp — Minimal read-only server with Web API support via pyzotero. No citekey resolution or in-document search.
• kaliaboi/mcp-zotero — Cloud-only (Zotero Web API). Metadata browsing, no fulltext.

To remove an existing Zotero MCP server before switching:

claude mcp remove zotero

Or delete the zotero entry from .mcp.json / claude_desktop_config.json / ~/.codex/config.toml manually.

Requirements

• Zotero 8+ with the local API enabled
• Python 3.11+
• Better BibTeX (optional but recommended)

License

MIT • Ulrich Atz (ulrichatz)