llm-for-zotero

llm-for-zotero is an open-source Zotero add-on by Yile Wang that adds an LLM assistant to the Zotero PDF reader and a library-wide chat. It supports provider APIs, local OpenAI-compatible models, browser-based WebChat, Codex App Server, and Claude Code. The documentation lists version 3.9.11 and support for Zotero 7, 8, and 9.

What It Is

The plugin is a research assistant embedded in Zotero. Open a PDF, click the LLM Assistant icon in the right-hand toolbar, and ask questions, summarize the paper, explain selected text, interpret figure screenshots, or compare several open papers. Answers are grounded in the paper content, and citations can jump back to the matching passage in Zotero. Conversations are saved locally and can be exported to Zotero notes in Markdown.

Agent Mode and Skills

Agent Mode (beta) is disabled by default. When enabled, the model can read and search the Zotero library, PDFs, and attachments, and can propose changes such as tags, metadata updates, collection moves, duplicate merges, and imports by DOI, arXiv ID, or URL. Write operations go through confirmation cards and can be undone. Plan Mode lets you review and approve a plan before anything is changed. It also offers file and shell tools, which require approval. Built-in actions include library audits, auto-tagging, and discovering related papers. Skills are Markdown guidance files with regex triggers that shape agent workflows; the docs describe 8 built-in skills and a portal for creating custom ones. A built-in MCP server lets external MCP-compatible agents use the same library tools.

Models and Backends

Preset providers include OpenAI, Gemini, Anthropic, MiniMax, GLM, DeepSeek, Grok, Qwen, Kimi, and GitHub Copilot, plus custom OpenAI-compatible endpoints such as Ollama, LM Studio, and vLLM. WebChat relays questions to ChatGPT, DeepSeek, or Gemini websites through the Sync for Zotero browser extension. Codex App Server and an experimental Claude Code bridge are also available. Claude Code mode does not yet support native Zotero API operations.

Parsing, Web Search, and Notes

Optional MinerU parsing (cloud or local mineru-api server) extracts Markdown with tables, equations, and figures, and caches it locally. For figure questions, the agent crops figures directly from the source PDF using MinerU as an index. General web search in Agent Mode uses Tavily, with source cards attached to answers. Notes can be saved to Zotero or as Markdown files in Obsidian, Logseq, or any local folder, using Pandoc citation syntax. A Usage Statistics panel summarizes local chat and token activity with CSV export.

Privacy and Data Flow

Data handling depends on the chosen backend: paper content goes to the configured provider, a local model endpoint, the WebChat site, or MinerU when parsing is enabled. Conversation history and cached paper context are stored locally. The plugin is licensed under AGPL v3.