PapersGPT
AI Research Agent for Zotero — Whole-Library Search, Literature Radar, Multi-PDF Chat & AutoPilot
PapersGPT is a professional AI research tool for Zotero, built for deep research across your entire library. More than chatting with a single PDF — go one step beyond: ask questions across your entire library, discover and track new research with Literature Radar, compare multiple papers side-by-side, and let AutoPilot batch-read collections overnight, with structured results saved directly into Zotero Notes.
It supports 30+ cloud LLMs (GPT-6.1, Claude 5.5, Gemini 4, DeepSeek v4.1, Grok…), built-in local models that run fully offline, or your own API key. Quick to start: a built-in trial of cloud AI chats — no API key needed.
🧭 PapersGPT at a Glance
| |
|---|
| What it is | Professional AI research agent and plugin for Zotero — deep-research your library: chat with PDFs, whole-library search, discover and track new papers |
| LLM support | 30+ cloud models (GPT-6.1, Claude 5.5, Gemini 4, DeepSeek v4.1, Grok, GLM 5.3, Kimi k3…) · built-in local models · Ollama · custom OpenAI-compatible APIs |
| Privacy | 100% local parsing, indexing and search; local LLMs run fully offline — your data never leaves your machine |
| Scale (tested) | 10,000 PDFs (42 GB, 170k pages) indexed in ~7 minutes; 19.5 ms average query time — benchmark details |
| Integrations | Zotero MCP Server (Claude Desktop, Claude Code, Cursor, ChatGPT desktop…) · Ollama · Agent SKILL |
| Compatibility | Zotero 7, 8, 9, 10 · macOS, Windows, Linux |
| License | AGPL-3.0 open source |
✨ Core Features
1. Whole-Library Search — ask across every paper in your Zotero library
PapersGPT indexes your entire Zotero library (thousands of PDFs). Just ask in natural language and get answers based on all your papers — not just the one you have open. Works at 10,000+ paper scale, and pairs well with multi-PDF chat for cross-paper analysis.
2. Literature Radar — discover & track new research
Subscribe to your research topics and Literature Radar monitors arXiv, PubMed, Semantic Scholar, Europe PMC, Crossref, bioRxiv, medRxiv for you. New papers arrive as a daily digest inside Zotero, visualized against your existing collections — new research finds you, instead of you chasing databases. Queries go straight from your machine to the source APIs, with no relay server.
3. Multi-PDF Chat & Comparison
Beyond chatting with a single PDF, you can chat with multiple PDFs at once: select several papers or an entire collection and compare methods, results and controversies side-by-side. Synthesize the cross-paper results into structured drafts or Zotero notes.
“Compare the sample sizes and p-values of all 50 clinical trials in my 'Cardiology' folder.”
4. AutoPilot — batch-read 100+ papers overnight
Set your research question once and AutoPilot autonomously reads the whole collection, extracting hypotheses, methods and findings into structured Zotero Notes while you sleep. Works with cloud models or fully local LLMs, protecting your privacy.
5. Verifiable Evidence — click-to-jump citations
Every AI response includes precise citations; clicking one takes you straight to the exact paper being cited. Accuracy you can verify, not just trust.
6. Zotero MCP Server & Agent Skills
Turn your local Zotero knowledge base into a primary "skill" for agentic platforms — Claude Desktop, Claude Code, Cursor, Windsurf, ChatGPT desktop and any MCP-compatible client. Setup takes just a few clicks from the PapersGPT settings panel; with mcp-zotero you can even index and search your whole library without launching Zotero. See MCP & SKILL Support below.
🔍 How It Works
PapersGPT parses and indexes every PDF in your Zotero library locally — structure-aware, so headings, sections and tables stay meaningful. When you ask a question, it retrieves the relevant passages across your library, synthesizes an answer with your chosen LLM (cloud or local), and attaches click-to-jump citations back to the exact source papers. Parsing, indexing and searching all happen on your machine — nothing is uploaded.
⚡ Performance at Scale
PapersGPT's engine parses document structure directly, so a library of 100 or 10,000+ PDFs stays fast to query — with 100% of parsing, indexing and searching happening locally, even in Airplane Mode.
| Library Size | Raw Data Size | Indexing Time | Avg. Query Time | Memory (RSS) | Index Size |
|---|
| 1,000 PDFs | 4.2 GB | 51.5s | 13.1 ms | 353 MB | 100 MB |
| 10,000 PDFs | 42 GB | 421s (7m 01s) | 19.5 ms | 2.21 GB | 901 MB |
👉 View Detailed Scalability Report & System Specs
🤖 Models: 30+ Cloud LLMs or Fully Local & Offline
- Cloud LLMs (30+): GPT-6, Claude 5.5, Gemini 4, DeepSeek v4.1, Grok, GLM 5.3, Kimi k3 and more — switch seamlessly, or plug in your own API key / any custom OpenAI-compatible endpoint.
- Built-in local models: Gemma 4, Qwen 3.5, DeepSeek Distill, Mistral 7B, Llama 3.1 — one-click download from Hugging Face, no complex installation. Run 100% on your machine, completely offline, no API keys needed.
Three ways to connect local LLMs:
- Integrated SOTA open-source LLMs — select a model from the "Local LLM" option and it is downloaded from Hugging Face and launched automatically.
- Custom OpenAI-compatible endpoint — connect to any privately deployed OpenAI-compatible LLM API.
- Ollama integration — use any LLM running in your local Ollama instance.
🔒 Privacy by design: parsing, indexing and searching all happen locally; local LLMs + local document processing ensure "zero-byte" data leakage from your device — critical for confidential research, unpublished work and institutional compliance.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
