llm-for-zotero: Deep Research Report on the Leading AI Plugin for Zotero

Forum topic · ✨步子哥 · 2026-06-14

Summary

llm-for-zotero is an open-source (AGPL v3) Zotero plugin by Yile Wang that embeds AI chat directly into the Zotero reader sidebar, aiming to eliminate context-switching between PDF readers and external chatbots. Written in TypeScript (96%), it supports Zotero 7/8/9, has ~1.9k stars, and ships frequent releases (v3.8.18, June 2026). Core features include full-paper Q&A with citation-linked answers, multi-paper comparison (up to 10 documents), multimodal inputs, Markdown note export with YAML frontmatter for Obsidian/Logseq, and MinerU-powered high-fidelity PDF parsing. A Beta Agent mode lets the LLM read and (with confirmation) write to the Zotero library via tools like query_library, apply_tags, and import_identifiers, with undo support. A Skill system offers 8 built-in Markdown-defined skills plus custom ones, and an integrated MCP server (http://localhost:23119/llm-for-zotero/mcp) exposes tools to external agents. The plugin supports six backend protocols: OpenAI/Gemini/DeepSeek APIs, local OpenAI-compatible endpoints (Ollama, LM Studio, vLLM), WebChat browser sync, Codex App Server, and experimental Claude Code bridging. Compared with rivals Aria (stagnant since Oct 2024) and Zotero-AI-Butler (batch automation focus), llm-for-zotero leads in feature depth, model flexibility, and modern architecture. Privacy-sensitive users can run fully local pipelines; main risks are Beta Agent stability and single-maintainer dependency.

This is an English translation of a Chinese technical deep-dive report on llm-for-zotero, published on zhichai.net.

Project Overview

| Attribute | Value | |---|---| | Repository | yilewang/llm-for-zotero | | Author | Yile Wang (@yilewang) | | License | AGPL v3 | | Languages | TypeScript 96%, CSS 2.8%, other 1.2% | | Community | ⭐ 1.9k, 🍴 96, 1297 commits | | Supported Zotero | 7, 8, 9 | | Latest release | v3.8.18 (2026-06-11) | | Open issues / PRs | 73 / 6 |

The author's stated philosophy: eliminate switching costs so AI becomes a natural extension of the reading experience. Traditional workflow (open PDF → paste into ChatGPT web → return to Zotero) is replaced by instant, in-context AI answers. The plugin is deliberately not a generic chat box, but an elegant tool that blends into Zotero — "concise but not shallow."

Core Features

Reading & Chat

Chat with the current PDF in Zotero; first query auto-loads the full text, subsequent queries use targeted retrieval
Citation-linked answers with jump-to-source in the original PDF
Multi-paper comparison: type / to reference up to 10 open papers
Full-text or method/results-specific summaries; select any passage for explanation
Multimodal: up to 10 figure screenshots; upload PDF/DOCX/PPTX/TXT/Markdown as context
Per-paper chat history, locally persisted, exportable to Markdown

Notes

Save answers/selections/conversations as Zotero built-in notes (Markdown supported)
File-based Markdown notes to any local directory — native fit for Obsidian and Logseq
Auto YAML frontmatter with Pandoc cite syntax [@citekey], compatible with the Zotero Integration plugin
MinerU-extracted figures can be embedded in notes

Agent Mode (Beta, off by default)

Read tools (direct execution): query_library, read_library, read_paper, search_paper, view_pdf_pages, search_literature_online (CrossRef, Semantic Scholar).

Write tools (human confirmation required): apply_tags, update_metadata, move_to_collection, manage_collections, manage_attachments, merge_items, trash_items, import_identifiers, import_local_files, edit_current_note, undo_last_action.

File system & scripts (confirmation required): file_io, run_command (shell), zotero_script (Zotero JavaScript).

Safety: undo for the last write action, up to 10 undo records per session. Cache-aware context: stable literature context and read evidence are retained; history is auto-compressed when the context window runs out.

Skills System

simple-paper-qa

evidence-based-qa

analyze-figures

compare-papers

library-analysis

literature-review

write-note

import-cited-reference

{ZoteroDataDir}/llm-for-zotero/skills/

MinerU PDF Parsing

mineru-api

pipeline

vlm

hybrid

MCP Server

http://localhost:23119/llm-for-zotero/mcp

Multi-Backend Support

| Backend | API key | Notes | |---|---|---| | Standard providers | Yes | OpenAI, Gemini, DeepSeek, Moonshot, etc. | | Local models | Usually no | Any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM) | | WebChat | No | Browser extension syncs ChatGPT/DeepSeek web sessions | | Codex App Server | No (ChatGPT Plus) | codex app-server runtime | | Claude Code | Claude Code auth | Experimental, local bridge |

Protocols: responses_api, openai_chat_compat, anthropic_messages, gemini_native, codex_responses, web_sync.

Competitor Comparison

| Project | Stars | Last update | Positioning | |---|---|---|---| | llm-for-zotero | 1.9k | 2026-06-11 | Deep AI research assistant + Agent mode | | Aria (ai-research-assistant) | 1.7k | 2024-10-20 | Lightweight embedded AI assistant | | Zotero-AI-Butler | 1.4k | 2026-06-05 | Automated note generation, batch processing | | Zotero-GPT | — | 404 / unmaintained | Early GPT plugin, defunct |

vs. Aria: llm-for-zotero wins on model breadth (Aria is GPT-4-only), local models, Agent mode, skills, MCP, multi-paper comparison, MinerU, WebChat, and update cadence. Aria only suits users wanting minimal OpenAI-only integration.

vs. Zotero-AI-Butler: Complementary rather than competing. Butler focuses on batch automated notes, mind maps, and one-click poster/summary generation; llm-for-zotero focuses on interactive deep reading, cross-paper dialogue, and Agent automation. Both can coexist.

Rankings per the report: feature richness / model breadth / modern architecture — llm-for-zotero > Zotero-AI-Butler > Aria; batch automation — Zotero-AI-Butler > llm-for-zotero > Aria; interactive chat — llm-for-zotero > Aria > Zotero-AI-Butler.

Code Quality

Based on windingwind/zotero-plugin-template; clean structure (addon/, src/, doc/, test/, typings/, zotero-plugin.config.ts)
Strengths: unified 6-protocol adapter layer, local caching, confirmation + undo safety, MCP standardization, bilingual docs
Risks: Beta Agent permission boundaries, Claude Code bridge dependency on third party, MinerU cloud/local service dependency, test coverage unclear

Community & Ecosystem

~30+ commits/month, releases every 1–2 weeks; contributors: @yilewang (lead), @jianghao-zhang (Codex/Claude Code), @boltma (file upload), @renyong18 (local MinerU)
Single-maintainer-dominated model: fast decisions but maintainer-bottleneck risk
Integrations: Obsidian/Logseq (mature), MinerU (mature), OpenAI ecosystem (mature), MCP (Beta), Claude Code (experimental)

Privacy Assessment

| Scenario | Risk | |---|---| | Cloud APIs / WebChat / cloud MinerU | Medium | | Local models + local MinerU + MCP (localhost) | Low — data never leaves the machine |

Limitations & Future Directions

Limitations: Zotero 7+ only; Agent mode Beta; experimental Claude Code; WebChat lacks Agent// commands; learning curve; maintainer bottleneck (high risk); AGPL v3 for commercial use.

Likely roadmap: Agent mode GA (~v4.0), Agent memory system (already on roadmap), native Claude Code support, cross-device sync (MinerU cache sync). Medium probability: more built-in skills, deeper MCP interoperability with other research AI tools. Risks: Zotero API changes, maintainer abandonment, feature bloat.

Recommendations by User Type

| User | Recommendation | |---|---| | Power users wanting a full AI research assistant | llm-for-zotero (first choice) | | Batch auto note generation | Zotero-AI-Butler | | Minimal setup | Aria (or llm-for-zotero WebChat mode) | | Privacy-critical, fully local | llm-for-zotero (local models + local MinerU) |

Key takeaways for similar tooling: multi-backend abstraction is essential; citation traceability is a core academic-AI requirement; Markdown-defined skills beat prompt templates for community contribution; MCP is the emerging interop standard; bilingual docs matter for reaching Chinese users.