Most PDF assistants work one file at a time; the reason to install this is asking a question across a whole collection and getting an answer whose citations you can click through to the source. Running it fully local — bundled model or Ollama — is a real option rather than a footnote, which matters if the library holds material you cannot upload.
A Zotero extension written in native C++ that indexes your library locally and answers questions across many documents at once. It ships a Zotero MCP server, so agent platforms such as Claude Code, Cursor and Windsurf can use that library as a source. Parsing, indexing and searching all happen on your machine and work offline.
- Questions answered across a single PDF, a selected set, a collection, or the entire library with "Search entire library" checked
- Cross-document comparison that keeps individual studies distinct — sample sizes and results pulled from many papers into one table
- Interactive citations on every answer: clicking one opens the specific paper it came from
- Local model options: bundled open-source models downloaded and run in one click, a custom OpenAI-compatible endpoint, or a local Ollama instance — alongside hosted models when you supply a key
- AutoPilot, which batch-processes a large set of papers against a research goal and writes the results back into Zotero notes
- A published SKILL file for OpenAI, Claude, Cursor, Claude Code, Codex and Gemini CLI, as an alternative to the MCP path
Zotero, plus the PapersGPT plugin installed from the released `.xpi` file. Then a model: a bundled local model downloaded from Hugging Face in one click, an OpenAI-compatible endpoint you configure, a local Ollama instance, or a hosted model with its API key. Indexing and search need no cloud service and run offline. Version 0.0.16. Professional capabilities are sold separately at papersgpt.com.