PapersGPT for Zotero: a local AI and MCP plugin for large PDF libraries

The problem PapersGPT targets: thousands of PDFs, no way to query them together

Zotero is good at storing references and bad at answering questions across them. The common workaround is to open one PDF, paste text into a chat window, and repeat. That breaks down as soon as the question spans documents: sample sizes across fifty trials, how a term changed meaning between 2015 and 2024, where two papers contradict each other. PapersGPT is aimed at that gap. The README frames it as "Industrial-Scale Multi-Document Intelligence for Zotero" and describes three scopes of query: a single open PDF, a multi-selection or a collection in the main Zotero window, and the entire library via a "Search entire library" checkbox. The intended user is a researcher with a library in the hundreds or thousands of PDFs who wants synthesis rather than summarization. The README also positions local indexing as the differentiator against tools that send document text to a service for embedding.

How the indexing and retrieval pipeline is put together

Two layers are visible in the repository. The extension itself is a Zotero add-on: addon/, src/, tsconfig.json and a scripts/ directory with build.js, start.js and stop.js, driven by npm scripts (build-dev, build-prod, start-z6, start-z7). That is a standard Zotero 6 and 7 plugin toolchain, and it is what produces the .xpi. The retrieval layer is where the claims live. The README says parsing, indexing and searching happen locally, with no cloud dependency, and that a structural index is built rather than embeddings: it describes the engine as moving "beyond the 'fuzzy' matching of standard semantic search by embeddings" and understanding document structure. The scalability table in the README lists 10,000 PDFs (42 GB) indexed in 421 seconds, average query time 19.5 ms, 2.21 GB RSS and a 901 MB index. Those numbers come from benchmark.md, which the README links as the detailed report; treat them as the project's own measurements on its own hardware, not as a general guarantee. The dependency list in package.json tells a more mixed story than "100% local": chromadb, @pinecone-database/pinecone, @dqbd/tiktoken and crypto-js are all present, so vector-store and tokenizer code ships with the plugin even though the README's headline is structural, local indexing. Answers are rendered with citations, and the README states that clicking a citation jumps to the specific paper.

Installing PapersGPT from the release .xpi and running a first query

The README gives a two-step quickstart. Download the .xpi from the release page, then install it into Zotero as a plugin. The download URL in the README points at the v1.2.0 asset, and the releases list shows papersgpt-v1.2.0 published on 2026-08-28. The README links to papersgpt.com/quickstart for the install details rather than spelling them out, so if the Zotero plugin dialog does not accept the file, that page is the place to check. There is no separate installer and no server to run for the default setup.

# The README points at this asset for v1.2.0
https://github.com/papersgpt/papersgpt-for-zotero/releases/download/papersgpt-v1.2.0/papersgpt-v1.2.0.xpi

Once installed, open a PDF in Zotero and launch the panel from the viewer toolbar, or use the keyboard shortcut the README documents.

macOS:   Command + Enter
Windows: Ctrl + Enter

For a multi-document query, select several items or a whole collection in the main Zotero window (Ctrl-click on Windows, Command-click on macOS) before launching the panel. For a library-wide query, tick "Search entire library". Then pick a model and supply its API key, which the README defers to papersgpt.com/models for the per-provider detail. The built-in prompts cover Summary, Background, literature review generation, Theoretical frameworks and Future directions. If you want no cloud at all, the README lists three local routes: one-click downloads of models it names as Gemma 4 12B and Qwen3.5 4B from Hugging Face, a custom OpenAI-compatible endpoint, or an existing Ollama instance. The README does not document rollback or an uninstall procedure.

Where the local-first claim gets complicated

The README's strongest sentence is "zero-byte" data leakage, and it is only true on the local-LLM path. Selecting a hosted model such as Claude, GPT or Gemini sends your prompts, and whatever document context the plugin attaches, to that provider. Indexing stays on disk; inference does not. The README is explicit that local models are one of three connection options, not the default, so a reader who skims the privacy line and then configures a cloud key has a different threat model than the one advertised. The second soft spot is the benchmark. A 7-minute index of 42 GB and 19.5 ms queries are the project's own figures, published in its own benchmark.md, with no third-party reproduction cited anywhere in the README. The numbers may well hold on the hardware described there; they are still a vendor measurement. Third, the repository's package.json carries version 0.0.16 while the releases are at v1.2.0, and its description still lists older models (DeepSeek-R1-Distill-Llama, QwQ-32B-Preview, Llama 3.2) that the README's headline no longer mentions. That mismatch suggests metadata lags the release train. Finally, the README's MCP section is thin: it says PapersGPT supports MCP and links to a separate repository for detail, so anyone planning to wire it into an agent host should expect to read that other project's documentation.

PapersGPT versus embedding-based Zotero AI plugins

The obvious comparison class is the other Zotero AI plugins that people search for by name, and the README draws the line itself: those tools use embedding-based semantic search, PapersGPT uses structural indexing and claims better cross-document accuracy as a result. The practical difference for a user is the failure mode. Embedding search returns passages that are semantically near your query, which is why it can surface a paragraph that reads relevant but answers nothing. A structural index, as described, is meant to keep document boundaries and sections intact so a comparison across fifty papers returns fifty attributable results rather than a fused blob. Whether that holds in practice is the thing to test on your own corpus, because the README's evidence is its own benchmark, not a head-to-head against a named alternative. The second axis is where the model runs. Plugins that assume a cloud API are simpler to set up and worse for sensitive or embargoed material; PapersGPT gives you the choice, at the cost of a heavier configuration surface (per-provider keys, an OpenAI-compatible endpoint field, or an Ollama address). The third axis is agent access. PapersGPT ships an MCP server and a SKILL definition so hosts like Claude Code, Cursor or Windsurf can query the library, which the embedding-based plugins in this space generally do not offer.

AutoPilot, MCP and the SKILL file: the batch and agent surface

Two features sit outside the chat panel. AutoPilot is described as an autonomous mode where you define a research goal and the plugin batch-processes over 1,000 papers, writing insights into Zotero Notes. The README calls it the first such tool for Zotero; that is a marketing claim, not a verifiable one, but the mechanism it implies (a queue of documents, a prompt template, output appended to notes) is plausible for an overnight run. The cost is that a batch job multiplies whatever you pay per token by the number of papers, so AutoPilot is the feature most sensitive to whether you are on a local model. The MCP server is the other one. It exposes the local knowledge base to agentic platforms, and the repository carries papersgpt-for-zotero/SKILL.md for agents that consume skills. Both of these are documented by reference rather than in full: the README sends MCP readers to a separate repository and OpenClaw users to clawhub.ai. If you are evaluating PapersGPT specifically as an MCP backend, the README alone will not tell you the tool schema, the transport, or which Zotero fields are exposed.

Licence, maintenance and what an upgrade costs you

PapersGPT is AGPL-3.0-or-later per package.json, and the repository ships a LICENSE file. For individual researchers this changes nothing day to day. For anyone embedding the plugin in a hosted product or a shared internal service, AGPL's network clause is the part to read with a lawyer, because it can extend source-disclosure obligations to users who interact with a modified version over a network. Nothing here is legal advice; the point is that AGPL is a real constraint on commercial redistribution, not a formality. On maintenance, the last push to the repository was on 2026-09-07, and the release history shows v1.0.0 on 2026-08-05, v1.1.0 on 2026-08-11 and v1.2.0 on 2026-08-28. That is a fast release cadence over roughly three weeks, which also means the upgrade path is the risk: three minor releases in a month with no migration notes in the README, and no documented rollback. Zotero plugins update through the add-on manager, and the repository carries update.json, update.rdf and update-template.json, so an update can arrive without you initiating it. If you pin a version for a lab machine, know that the plugin's own updater metadata is what decides otherwise.

Editorial conclusion

Adopt PapersGPT if you already live in Zotero and your library is large enough that per-file chat has stopped being useful, and if you accept AGPL-3.0 terms and the cloud exposure that comes with hosted model APIs. Do not adopt it if you need a documented, stable API surface or a plugin that runs without any third-party model account. Verify first that your Zotero version matches the current release, that your chosen model is reachable from your network, and that the indexing behavior on your own PDFs matches the numbers in benchmark.md before you point it at a 10,000-file library.

Frequently asked questions

Official sources

1. License: AGPL-3.0
2. papersgpt/papersgpt-for-zotero on GitHub
3. Project website
4. README
5. Releases