Upgrade Storage[Plugin] Zocrates – Open-Source Agentic Research Assistant
Hi everyone,
Like many of you, I have several Zotero libraries—some with thousands of papers—representing years of careful curation. I wanted a plugin that could surface connections across a sprawling library and synthesize information as reliably, transparently, and with as much rigor as possible.
As an AI engineer, I wanted to move beyond typical "chat-with-PDF" tools that chop papers into arbitrary fragments and hallucinate citations. The result is Zocrates, an open-source research assistant built around structured document reading, auditable attribution, and modern agentic workflows.
Under the hood, Zocrates first transforms your PDFs, web snapshots, and Zotero notes into clean, self-contained Markdown knowledge bundles (preserving document structure, tables, and figures). From there, you can use Zocrates in three ways:
- As a native Zotero companion: Adds a sidebar inside Zotero with an agentic workflow tailored for academic search and synthesis. You can chat with your library or collections, send text excerpts and figure/table crops directly from Zotero's PDF reader into the chat, save syntheses back as native Zotero notes, and search online databases (OpenAlex/Crossref/Unpaywall) with one-click PDF import directly into Zotero.
- Direct folder access for coding agents: Point your favorite CLI/coding agent (Claude Code, Codex, Pi, etc.) directly at the folder of converted Markdown bundles and let it navigate them. This works especially well for focused collections (up to a few hundred papers).
- Via MCP (Model Context Protocol): Connect external AI apps (Claude Desktop, Antigravity, etc.) to your converted library and Zocrates' search, targeted document reading, and discovery tools.
Using the native Zotero plugin provides the richest experience, including:
- Auditable citations & claim verification: Responses include inline citations with confidence scores, verbatim supporting passages, one-click jumping to the exact page in Zotero's PDF reader, and optional sentence-level attribution.
- Offline paper relations graph: Automatically builds and visualizes a network of relationships across your library (100% locally, without LLM calls) that both you and the agent can use to explore connected literature.
- Full user control:
- Privacy: Everything (indices, converted bundles, and chat history) stays on your machine, and you can run the entire stack locally with Ollama or vLLM if you have the hardware.
- Choice: Bring your own local models or cloud API key (OpenRouter), and even mix-and-match local and cloud models per agent role.
- Pricing: Free and open-source (licensed under AGPL-3.0, just like Zotero). Your converted data is portable and yours forever. No subscriptions, paywalls, or platform lock-in. You only pay your chosen model provider's published API rates (or nothing at all if running locally or using free-tier models).
Compatibility & Installation Expectations:
- Zotero Version & OS: Targets Zotero 10+ on Windows 10+, macOS, and Linux.
- Zero-config backend setup: Installs as a standard Zotero add-on (.xpi). On first launch, Zocrates automatically sets up an isolated Python environment and local backend inside your Zotero data folder without touching your system Python.
- System resources: Expect ~2 GB of free disk space for the local environment, included local embedding model weights, and search indices (8 GB+ RAM recommended; 16 GB+ if running local LLMs).
Links & Getting Started:
- Project Overview: https://alexplaka.github.io/Zocrates/
- GitHub Repository & Releases: https://github.com/alexplaka/Zocrates
- Documentation & Setup Guide: https://alexplaka.github.io/Zocrates/docs
Building cross-platform software that runs locally across different machines and hardware setups has its challenges. I am working toward a v1.0 release in the coming weeks. If you try it out, please report any bugs or rough edges on the GitHub repo Issues page (or using the built-in "Report Issue" button inside the plugin)!
If you'd like to get involved as a contributor or help shape its evolution, please reach out. I am especially interested in hearing about the research workflows you'd most like an agentic companion to support.
Thank you for testing, and happy researching!
Alex Plakantonakis
Like many of you, I have several Zotero libraries—some with thousands of papers—representing years of careful curation. I wanted a plugin that could surface connections across a sprawling library and synthesize information as reliably, transparently, and with as much rigor as possible.
As an AI engineer, I wanted to move beyond typical "chat-with-PDF" tools that chop papers into arbitrary fragments and hallucinate citations. The result is Zocrates, an open-source research assistant built around structured document reading, auditable attribution, and modern agentic workflows.
Under the hood, Zocrates first transforms your PDFs, web snapshots, and Zotero notes into clean, self-contained Markdown knowledge bundles (preserving document structure, tables, and figures). From there, you can use Zocrates in three ways:
- As a native Zotero companion: Adds a sidebar inside Zotero with an agentic workflow tailored for academic search and synthesis. You can chat with your library or collections, send text excerpts and figure/table crops directly from Zotero's PDF reader into the chat, save syntheses back as native Zotero notes, and search online databases (OpenAlex/Crossref/Unpaywall) with one-click PDF import directly into Zotero.
- Direct folder access for coding agents: Point your favorite CLI/coding agent (Claude Code, Codex, Pi, etc.) directly at the folder of converted Markdown bundles and let it navigate them. This works especially well for focused collections (up to a few hundred papers).
- Via MCP (Model Context Protocol): Connect external AI apps (Claude Desktop, Antigravity, etc.) to your converted library and Zocrates' search, targeted document reading, and discovery tools.
Using the native Zotero plugin provides the richest experience, including:
- Auditable citations & claim verification: Responses include inline citations with confidence scores, verbatim supporting passages, one-click jumping to the exact page in Zotero's PDF reader, and optional sentence-level attribution.
- Offline paper relations graph: Automatically builds and visualizes a network of relationships across your library (100% locally, without LLM calls) that both you and the agent can use to explore connected literature.
- Full user control:
- Privacy: Everything (indices, converted bundles, and chat history) stays on your machine, and you can run the entire stack locally with Ollama or vLLM if you have the hardware.
- Choice: Bring your own local models or cloud API key (OpenRouter), and even mix-and-match local and cloud models per agent role.
- Pricing: Free and open-source (licensed under AGPL-3.0, just like Zotero). Your converted data is portable and yours forever. No subscriptions, paywalls, or platform lock-in. You only pay your chosen model provider's published API rates (or nothing at all if running locally or using free-tier models).
Compatibility & Installation Expectations:
- Zotero Version & OS: Targets Zotero 10+ on Windows 10+, macOS, and Linux.
- Zero-config backend setup: Installs as a standard Zotero add-on (.xpi). On first launch, Zocrates automatically sets up an isolated Python environment and local backend inside your Zotero data folder without touching your system Python.
- System resources: Expect ~2 GB of free disk space for the local environment, included local embedding model weights, and search indices (8 GB+ RAM recommended; 16 GB+ if running local LLMs).
Links & Getting Started:
- Project Overview: https://alexplaka.github.io/Zocrates/
- GitHub Repository & Releases: https://github.com/alexplaka/Zocrates
- Documentation & Setup Guide: https://alexplaka.github.io/Zocrates/docs
Building cross-platform software that runs locally across different machines and hardware setups has its challenges. I am working toward a v1.0 release in the coming weeks. If you try it out, please report any bugs or rough edges on the GitHub repo Issues page (or using the built-in "Report Issue" button inside the plugin)!
If you'd like to get involved as a contributor or help shape its evolution, please reach out. I am especially interested in hearing about the research workflows you'd most like an agentic companion to support.
Thank you for testing, and happy researching!
Alex Plakantonakis