What it does

Zotero-AI-Butler is a plugin that intercepts your PDFs before they rot in your “read later” pile. It sends papers to your choice of LLM—Gemini, OpenAI, Anthropic, Volcano Ark, Ollama, or any OpenAI-compatible endpoint—and writes the resulting summaries directly into Zotero as Markdown notes. You can trigger it manually, let it auto-scan new arrivals, or batch-process your backlog. It also generates mind maps, multi-paper literature reviews, and (in v3.0) “academic poster” images via Nano Banana Pro.

The interesting bit

The plugin treats Zotero as the source of truth, not just a storage bin. Notes live inside the item they describe, follow Zotero’s native interaction patterns, and can be pinned as a persistent sidebar so switching papers doesn’t break your flow. The “producer-consumer” task queue with configurable rate-limiting is a nice touch for anyone who’s hit API caps.

Key highlights

• Multi-modal PDF ingestion: Base64-encodes PDFs for vision-capable models (Gemini 2.5 Pro recommended), or falls back to text extraction for models that can’t “see”
• Three trigger modes: right-click immediate, auto-cruise on new drops, or bulk backlog scan
• Customizable prompts: all templates are editable, with variable substitution ({{title}}, {{authors}}, etc.) and live preview
• Built-in追问 (follow-up): sidebar supports multi-turn Q&A against the generated summary, with LaTeX rendering
• No proxy server: requests go straight from your machine to your configured API; the plugin claims it touches no data itself

Caveats

• Bring your own API keys—and your own rate-limit anxiety. The plugin is free but the LLM bills are yours
• “Pre-release” label on the对话与追问 feature; stability unclear
• Auto-scan is off by default to avoid performance hits, so “set and forget” requires one manual toggle
• The README’s recommended “free” Gemini 3 Pro access involves deploying a separate proxy project (gcli2api), which is not exactly turnkey

Verdict

Grad students and researchers drowning in unread PDFs should take a look—especially if you already pay for an LLM API. If you expect a polished consumer product with bundled AI, this is not it; it’s a power-user’s glue code with a surprisingly thoughtful queue system.