yilewang/llm-for-zotero

llm‑for‑zotero – an LLM‑powered research assistant inside Zotero

What it is

• A Zotero add‑on (distributed as a .xpi file) that embeds a chat interface right in the Zotero PDF reader.
• The chat is backed by any LLM you can reach via an API key, a local OpenAI‑compatible server, ChatGPT WebChat, the Codex App Server, or Claude Code.
• It lets you ask questions about the currently opened paper, summarize sections, analyse figures, compare multiple papers, and even run library‑wide actions (tagging, importing, note‑writing) through an Agent Mode.

Key capabilities

Feature	What you can do
Paper‑level chat	Ask the model to summarise a whole PDF, explain a selected paragraph, or interpret a screenshot of a figure.
Grounded citations	Answers include clickable citations that jump back to the exact passage in Zotero.
Multi‑paper context	Use the / command to pull in other open papers as additional context, enabling side‑by‑side comparison.
File‑based notes	Export generated notes to any local Markdown folder (Obsidian, Logseq, plain .md), with proper YAML front‑matter and Pandoc‑style citation keys.
Agent Mode (beta)	The LLM can read/write the whole library: search items, update metadata/tags, import new records, create or edit notes, and run custom skills that encode reusable workflows. All write actions are confirmed and can be undone.
Web search	Via the Tavily service the agent can browse the public web, read pages, and attach source links to its answers.
Custom providers	Supports OpenAI‑style chat, Anthropic messages, Gemini native, as well as the experimental Claude Code bridge and Codex App Server.
MinerU PDF parsing	Optional higher‑fidelity extraction for tables, equations, and figures, with a local mineru‑api server.
Standalone window	Open the assistant in its own window with tabs for paper chat, library chat, and conversation history.

How to install

1. Download the latest .xpi from the Releases page.
2. In Zotero → Tools → Add‑ons → gear icon → Install Add‑on From File and select the file.
3. Restart Zotero.
4. Open Preferences → llm‑for‑zotero, pick a provider, fill in the base URL, API key and model name, then hit Test Connection.
5. Open any PDF and click the LLM Assistant icon on the right‑hand toolbar to start chatting.

Typical workflow

• Open a PDF → ask “Summarise the methods section”.
• Click a generated citation to jump to the source passage.
• Use / to add another paper as context, then ask “How do the results differ?”.
• When you’re ready, say “Save this summary to Obsidian” and the plugin writes a Markdown note to the configured vault.
• Enable Agent Mode to let the model tag the paper, import related works from OpenAlex, or edit the Zotero metadata after you confirm each step.

Configuration highlights

• Provider panel – choose among several protocols (responses_api, openai_chat_compat, anthropic_messages, gemini_native). You can set different models for different tasks (e.g., a multimodal model for figure analysis, a text‑only model for summarisation).
• Notes Directory – give a nickname (e.g., Obsidian) and a path; the agent will write notes there using Pandoc citation syntax ([@citekey]).
• Skills – a built‑in library of eight task‑specific instruction files (evidence‑based QA, figure analysis, literature review, etc.). You can edit or add new skills via the Skills portal.
• Web search – supply a Tavily API key to enable the web_search/web_read tools.

Privacy & data flow All content sent to the LLM is limited to what you explicitly include (the current paper, selected text, uploaded files, or web pages you ask to read). The plugin does not automatically upload your entire Zotero library.

Roadmap (as listed)

• Further polish of Claude Code integration.
• More built‑in skills and a skill‑creation UI.
• Enhanced cache handling for very long research sessions.
• Additional provider support and multimodal model options.

Who might find it useful? Researchers who already use Zotero for reference management and want an on‑demand AI assistant for literature review, note‑taking, and library organisation without leaving the Zotero environment.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.