Your Zotero library represents years of careful curation—hundreds or thousands of papers, annotations, and personal notes. Zocrates pairs state-of-the-art multi-agent AI with uncompromising scholarly rigor directly on your desktop.
Zocrates
Coordinator + ReAct Pipeline
Projecting queries, keys, and values into h = 8 lower-dimensional subspaces (dk = dv = dmodel/h = 64) allows the model to jointly attend to distinct representation subspaces at different positions without averaging inhibition .
While standard multi-head attention materializes the full N × N intermediate attention matrix in GPU memory, IO-aware tiling partitions inputs into fast on-chip SRAM blocks , while sparse attention combines local sliding windows with global tokens for linear 𝒪(N) sequence scaling .
"Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this."
Rather than locking your curated library behind closed proprietary clouds or treating academic PDFs as opaque text dumps, Zocrates respects how scholars read, verify, and think.
The name Zocrates is a portmanteau of Zotero and Socrates, honoring the classical philosopher renowned for dialectical inquiry and intellectual honesty. Rather than acting as an ungrounded oracle, Zocrates embodies the Socratic method by actively questioning literature, probing research gaps, and demanding verifiable evidence for every claim. It transforms your personal Zotero collection from a passive repository into an inquisitive, rigorous dialogue partner.
Use Zocrates as an embedded split-pane companion inside Zotero, as an open Markdown knowledge bundle engine on your filesystem, or as a local Model Context Protocol (MCP) server for external assistants like Claude Desktop, Cursor, Pi & others (any application that supports MCP).
Your search indices, converted Markdown bundles, strictly local deterministic paper relation graphs, and chat histories live 100% on your machine—fortified by 256-bit loopback auth, strict CORS isolation, safe DOM sanitization, and encrypted credential storage.
Every synthesized claim is anchored in exact document passages by default, verified per-occurrence by calibrated AI decision models, drilled down to sentence-level micro-attributions on demand, and backed by Two-Tier Source reporting so zero evaluated literature is ever silently dropped.
From native LaTeX mathematics, inline extracted paper figures, and user-curated long-term memory via native Zotero notes to autonomous online literature discovery across OpenAlex, Crossref, and Unpaywall with one-click Zotero collection import.
Every researcher has a distinct technical setup. Switch between modes below—or use all three simultaneously against the same local knowledge base.
Designed for researchers who want a deeply integrated desktop workspace without ever leaving Zotero's window.
read_document(asset_id='sX')) for targeted investigations—preserving complete arguments, unbroken mathematical proofs, and tables without chunk salad fragmentation.1. Your Research Question & Context • Focus on specific papers, collections, or your whole library • Right-click text selections or crop figures and formulas in Zotero's reader 2. In-Depth Literature Investigation • Reads paper structures, section headings, and full text • Follows thematic links and connected literature across your library • Extracts key methodology, evidence, and substantive arguments 3. Rigorous Claim Verification & Audit Table • Passage-level grounding by default via calibrated AI decision models • On-demand sentence-level micro-attribution highlighting exact supporting text • Only attaches citation pills to claims backed by source text (with compound prompt support) • Explore every claim, probability score, and quote in the Claim Verification Inspector • Accounts for all consulted papers with complete transparency 4. Scholar-Ready Synthesis • Interactive Author-Year citation pills with live Evidence Peeks • Clean mathematical formulas and inline paper figures • Anticipates your next research questions with smart suggestions • 1-click save to native Zotero notes for cross-session searchable memory
For researchers who love plain text, Unix philosophy, and direct filesystem ownership. Point Cursor, Claude Code, Pi, or other favorite coding agent straight at your corpus/ directory—zero external bridges required.
images/), and tables (tables/).index.md catalog and AGENTS.md structural guide so any coding agent understands your library immediately.# Zocrates OKF v0.2 Navigation Protocol for Coding Agents 1. Inspect `index.md` to locate relevant attachment keys. 2. Read `<item_key>/toc.md` or `toc.json` first for section line offsets (s0..sN) & figure/table catalogs (f1..fN, t1..tN). 3. Read targeted line ranges in `<item_key>/document.md` and traverse `item_relations.md` for 1-hop corroborated papers.
Give external AI coding tools and desktop assistants (Claude Desktop, Cursor, Pi or any other agentic application that supports the open Model Context Protocol) precision access to your entire Zotero library while you write code or draft manuscripts.
python -m zocrates.mcp), allowing external AI assistants to talk to Zocrates with zero data duplication.{
"mcpServers": {
"zocrates": {
"command": "python",
"args": ["-m", "zocrates.mcp"]
}
}
}
── What This Enables in Claude, Cursor, Pi & Other MCP Assistants ──
• Semantic & keyword search across your entire Zotero library
• Inspect paper outlines, section headings, figures, and tables
• Read targeted excerpts and full converted Markdown documents
• Traverse verified citations and thematic paper relationships
• Dedicated literature review prompts and library indexes
Zocrates moves beyond static chat wrappers. It operates as an autonomous research agent that investigates your whole library, queries global scholarly databases, pauses at critical milestones for your guidance, and intelligently manages long investigation context.
Standard AI search tools chop papers into arbitrary fragments, scattering sentences, breaking equations, and severing tables. Zocrates uses an autonomous multi-agent retrieval pipeline combining broad semantic search across your entire library with targeted, vectorless reading: the search agent inspects the author's Table of Contents and reads entire cohesive sections (read_document(asset_id='sX')), preserving complete arguments and derivations with maximum token efficiency.
Queries scholarly databases, deduplicates against your local Zotero library (via DOI, arXiv ID, and title Levenshtein similarity), stages verified Open Access PDFs, and writes directly into your target collection.
Autonomous research agents should never be uncontrollable black boxes. When exploring complex topics or comparing literature, Zocrates investigates deeply and pauses at natural milestones with a clear checkpoint summary of what has been discovered and what remains open. You decide whether to extend the search with custom guidance or synthesize the findings right away.
Extended multi-turn literature reviews quickly hit model context limits or inflate token costs. Running in the background with zero latency impact, an AI decision model evaluates topic closure to compact earlier turns while retaining the latest completed turns 100% verbatim. Raw search loops and chunk payloads are consolidated, keeping multi-hour research conversations fast, affordable, and context-rich.
Save individual assistant answers, comparative syntheses, or entire research sessions directly into Zotero as native rich-text notes with a single click. Because Zocrates converts and indexes your Zotero notes alongside PDFs into its local vector and full-text engines, your past discoveries automatically become searchable in future sessions.
Unlike closed cloud systems where session knowledge evaporates when a chat is archived, your research memory stays in your personal Zotero database—portable, backed up, and permanently yours.
Every subsystem in Zocrates is built to eliminate hallucinations, enforce verified attribution, surface visual and mathematical nuance, and connect literature across your collection.
Every synthesized claim is anchored in exact source passages by default. Fast, calibrated AI decision models evaluate claims on a per-occurrence basis ($p \ge 0.60$), automatically pruning unsupported citations and compound brackets so citations never hallucinate. Click any in-text pill to inspect exact section breadcrumbs, verbatim source quotes, and calibrated confidence badges:
While passage-level grounding substantiates the broader paragraph, deep scholarship often requires verifying exact numbers, parameters, and mechanisms. Click "🛡️ Claim Verification Inspector" on any response to open a standalone tabular audit. Inside, on-demand sentence-level micro-attribution isolates the exact supporting sentences in-place and flags any missing factual assertions without cluttering your chat:
Generated strictly locally using deterministic algorithms (statistical keyphrase extraction and reciprocal rank fusion hybrid scoring)—requiring zero LLM calls and zero cloud requests. Standard graph tools force every paper to connect to a fixed number of neighbors, hallucinating links between unrelated topics. Zocrates forms edges only when dense semantic vectors and statistical keyphrases corroborate each other mathematically. Click any node in the live graph below:
Anchor conversations directly to specific passages, figures, or equations while reading in Zotero's PDF reader. Right-click any highlighted text or draw an area annotation box around a diagram or chart to add it directly to your chat context. Visual crops are passed natively to vision-capable models, with automatic parent paper linking and live annotation resize tracking.
Academic communication demands more than plain text. Zocrates natively renders inline and block LaTeX equations via KaTeX, displays high-resolution figures extracted directly from your papers inside chat responses, and enriches uncaptioned diagrams with Vision-Language Model <synthetic_caption> blocks indexed in both ChromaDB and SQLite FTS5.
Complex taxonomies, biological pathways, and system architectures render as 100% offline Mermaid diagrams with a non-destructive toggle to copy the raw diagram code into your manuscript.
Unpublished manuscripts and proprietary datasets belong on your workstation. Zocrates keeps every database and converted bundle on your disk while giving you pragmatic choice over inference compute.
Running state-of-the-art reasoning models locally requires heavy GPUs with generous VRAM. For most researchers, connecting to OpenRouter delivers the best balance of frontier intelligence, speed, and zero hardware strain:
Working with strictly confidential clinical data, embargoed patents, or air-gapped laboratory machines? Run Zocrates completely disconnected from the internet:
You are never locked into an all-or-nothing choice between cloud and local. Zocrates allows you to seamlessly mix and match models and providers across different layers of your research workflow:
Run fast, free local models (via Ollama or vLLM) for conversational chat and multi-step retrieval loops, while delegating final literature synthesis to frontier cloud reasoning models.
Keep your library vector embeddings 100% offline with hardware-accelerated Jina v5, run citation verification through local Laya, and connect chat to OpenRouter—in any combination tailored to your hardware and privacy needs.
Keep raw library navigation and intermediate tool queries strictly local on your workstation. Only the distilled, highly relevant evidence excerpts are forwarded to the synthesis model.
Your entire research brain travels with you. If you work across multiple workstations (such as a Linux desktop in the lab and a MacBook at home), you can simply copy your zocrates-data folder:
All databases, Markdown bundles, and image extractions use relative file paths—never hardcoded computer paths like /home/user or C:\Users. Links never break across machines or operating systems.
When launching on a new computer, SHA-256 content verification immediately recognizes your synced PDFs. Vector embeddings, search indices, and literature graphs start up instantly with zero duplicate processing.
All multi-turn chat sessions, active attachment context excerpts, and paper relation networks remain completely intact. Pick up your research conversations exactly where you left off.
The assistant connects strictly to your local Zotero app, blocking outside websites and unauthorized programs from accessing it.
Standard browser isolation prevents regular web pages you visit from ever connecting to your Zocrates assistant or reading your data.
Cleans notes and paper text before display so malicious scripts or harmful formatting can never run on your workstation.
Confines all document reading strictly to your designated library folders, safeguarding the rest of your computer.
API keys are never stored in plain text—they are safely encrypted using Mozilla's built-in secure credential storage.
No manual Python environment setup or command-line scripts required. Zocrates installs as a standard Zotero 10+ extension and provisions its hermetic runtime automatically.
Download the latest zocrates.xpi release file from the official GitHub Releases page. In Zotero 10+, open Tools → Plugins, click the gear icon ⚙️, and select Install Plugin From File…
On first launch, Zocrates automatically provisions an isolated, per-profile runtime under <zotero-data>/zocrates-data/ with pinned dependencies—and shuts down cleanly whenever Zotero closes.
Open Zotero Settings → Zocrates. Enter your OpenRouter API key (recommended for instant frontier models) or point to a local Ollama / vLLM server. Tailor your persistent User Research Profile and start chatting!
Zocrates does not replace Zotero—it unlocks the full intellectual potential of the library you have already built.
| Dimension | Native Zotero Experience | With Zocrates Enabled |
|---|---|---|
| Ways to Work | Manual reading across tabs and static collection folders. | Three flexible modes: Native Zotero companion sidebar, open Markdown conversion engine (OKF v0.2), and local MCP server for external coding agents. |
| Data Privacy & Security | Local reference database on your computer; plaintext preference storage. | 100% local-first data with defense-in-depth security: Search indices, Markdown bundles, relation graphs, and chat histories live on your machine—fortified by 256-bit loopback auth, strict CORS, allowlist DOM sanitization, path containment, and encrypted Login Manager key storage. |
| PDF Reading & Excerpts | Highlighting and annotating text stored as static marks in PDFs. | Interactive context excerpts: Right-click highlighted text or draw area boxes around figures and equations in Zotero's PDF reader to analyze them with multimodal AI. |
| Scholarly Presentation | Plain unformatted text in notes or basic chat extensions. | Rich native rendering: LaTeX mathematical expressions, inline high-res figures from your papers, interactive Mermaid diagrams (with code toggle), and syntax-highlighted code blocks. |
| Online Literature Discovery | Manual searching on external websites, downloading PDFs, and dragging into Zotero. | Agentic online search & one-click import: Queries OpenAlex, Crossref, and Unpaywall, deduplicates against your local Zotero DB, stages verified Open Access PDFs, and imports directly into collections. |
| Evidence & Attribution | Manually searching open PDFs to verify assertions. | Two-tier attribution & Claim Verification Inspector: Passage-level grounding by default via calibrated AI decision models, fine-grained sentence micro-attribution on demand, compound bracket pruning, standalone Claim Verification Inspector window with probability scores ($p$) and verbatim source evidence, exact section breadcrumbs, and 1-click reader jump. |
| Attribution Completeness | External AI search tools silently drop retrieved sources they don't cite. | Complete Source Transparency (Zero Silently Dropped Sources): Transparently reports all direct evidence in Cited Sources and every other paper consulted during retrieval in an expandable Other Consulted Sources list. |
| Literature Retrieval | Basic keyword search in metadata or raw PDF text. | Multi-step literature search: Combines semantic search, section-by-section outline navigation, and targeted passage reranking to find and synthesize the most relevant evidence across your library. |
| Deep Research Investigation | Fixed limits that stop prematurely or run up unpredictable costs. | Interactive research checkpoints: For complex investigations, Zocrates summarizes progress at major milestones, highlights open questions, and lets you choose whether to extend the search or synthesize findings immediately. |
| Document Understanding | Unindexed diagrams and dense two-column formulas locked in PDFs. | Multi-engine conversion & visual descriptions: Layout-aware parsing, synthetic figure descriptions for search indexing, structured paper outlines, and first-class Zotero notes. |
| Literature Relationships | Manually linking related items one by one in metadata. | Strictly local deterministic relation graph: Generated 100% locally using deterministic algorithms (statistical keyphrases and hybrid retrieval with zero LLM calls). Corroborates semantic meaning with tags in an offline Canvas 2D visualizer. |
| Cross-Tool Extensibility | Data accessible only inside the Zotero desktop window. | Local Model Context Protocol (MCP): Lets Claude Desktop, Cursor, Pi & other MCP-compatible assistants search and read your library with zero duplicate memory. |
| Adaptive Context Compaction | Stateless reference manager or unmanaged context that exceeds token limits and slows down. | Autonomous post-turn context compaction: Intelligent rolling summarization consolidates earlier conversational turns in the background while keeping recent turns verbatim, thus keeping marathon research sessions fast, responsive, and within model token budgets. |
| User-Curated Long-Term Memory | Chat insights lost upon closing the window; notes disconnected from AI tools. | Native Zotero Note Memory: Save individual answers or entire sessions as native Zotero notes. Notes are automatically converted and indexed into your local search engine, making previous research permanently searchable in future sessions. |
| Cross-Device Portability | Syncs PDFs and metadata, but external AI tools require cloud re-uploads or hours of re-indexing on every new workstation. | Instant Copy-Paste Portability: Simply copy the zocrates-data directory to another synced computer (Linux, macOS, Windows). Fully portable relative links and content verification guarantee zero re-indexing, zero re-conversion, and instant session continuity out of the box. |