DocsAgent MCP — Zotero MCP Server 📚⚡

DocsAgent gives AI agents instant, private access to your personal knowledge base.
@docsagent/mcp-zotero is the spec-driven MCP (Model Context Protocol) server (Zotero is
the first supported source) that lets any AI agent — Claude Desktop, Cursor, Cline, Qwen Code, or
any MCP client — search, read, and write your Zotero library through a resident C++ search
engine. BM25 full-text search + query-ranked passage retrieval over 1,000+ PDFs at
~15 ms, fully local (RAG-ready knowledge base).

• 🔒 Local-first & private — the engine reads your Zotero library directly on your
  machine. Your PDFs never leave it.
• ⚡ Native C++ search core — inverted-index BM25 + passage ranking, millisecond lookup,
  low memory footprint (160–227 MB for a 1,500-paper library).
• 🧩 8 MCP tools — 5 read + 3 write, with JSON-schema validated arguments, token budgets,
  result dedup, and a three-layer write safety gate.
• 🌐 Two transports — stdio for local MCP clients, Streamable HTTP for remote deployment
  (origin checks, API-key / OAuth 2.0 token introspection, per-request RBAC, /health probe).
• 🐍 Two shells, one core — this TypeScript package and a feature-equal
  Python wrapper ship the same tools over the same JSON-RPC contract.

Architecture

MCP Client (Claude Desktop / Cursor / Cline / Qwen Code / any MCP host)
        │  stdio (local)   or   Streamable HTTP  /mcp  (remote)
        ▼
MCP shell  ← this package (@docsagent/mcp-zotero / docsagent-mcp-zotero)
   · tool schemas (spec-driven), argument validation, token budget, dedup
   · write orchestration via the Zotero local API, write safety gate, RBAC
   · group-library sync via the Zotero Web API
        │  JSON-RPC 2.0 over HTTP ({coreHost}:{httpPort}/rpc, cpp-httplib)
        ▼
DocsAgent Core (resident C++ engine, papersgpt-agent)
   · reads ~/Zotero/zotero.sqlite + storage/ directly on your machine
   · builds & serves the full-text index (BM25 + passage ranking)

The shell never spawns the core during tool calls and never touches your Zotero files.
The core runs as a background service and stays available across MCP client restarts.
Tool schemas and error codes: spec/.

Quick Start

1. Start the core

npx @docsagent/mcp-zotero start           # spawn the bundled core for your platform
npx @docsagent/mcp-zotero status          # pid / endpoint / version

Python shell (same verbs, under the core subcommand):

pip install ./python                      # build the wheel locally (PyPI upload pending)
docsagent-mcp-zotero core start

(core stop / core restart also available. The core reads your Zotero data directory,
builds the full-text index, and serves JSON-RPC on http://0.0.0.0:23120/rpc.)

2. Configure your MCP client

Claude Desktop / Cursor / Cline / Qwen Code (mcpServers):

{
  "mcpServers": {
    "docsagent-zotero": {
      "command": "npx",
      "args": ["-y", "@docsagent/mcp-zotero"]
    }
  }
}

Python shell (same tools, same contract, installed from this repo) — mcpServers:

{
  "mcpServers": {
    "docsagent-zotero-py": {
      "command": "docsagent-mcp-zotero"
    }
  }
}

Install the Python package first (step 1) so docsagent-mcp-zotero is on your PATH.

On startup the shell connects to the core, loads sources, and checks index status. If the
core is not running it fails fast with startup instructions — it never spawns anything.

MCP Tools (external API)

8 tools, 5 read + 3 write. Schemas are the single-sourced contract in
spec/tools/*.json (mirrored into both packages); arguments are validated
before handlers run and failures map to typed docsagent error codes.

list_sources

Every searchable source with capabilities, supported targets/includes/browse modes,
filters, and document counts. Call this first.

search

Cross-entry search over the whole library.

Returns results[] with global ids (zotero:KEY), titles, relevance, snippets; multi-target
searches group by target. Results are deduped (id, then normalized title + year) and packed
under a token budget.

get_content

Read one entry. mode=passages (query-ranked passages, k) or mode=fulltext
(offset pagination with nextOffset). Notes return their body with tags and metadata.

get_metadata

include: metadata, abstract, annotations, notes, citation (bibtex / csljson /
formatted via citationFormat/citationStyle). Notes are packed under the token budget.

list_library

Browse modes: collections (drill-down via parentId), items (by containerId),
tags, saved_searches, standalone_notes.

Write tools (three-layer safety gate)

Write safety gate (spec/algorithms/write-gate.md): layer 1 write tools are not
registered unless enableWrites=true; layer 2 confirmed=false returns a preview and
consumes no rate-limit quota; layer 3 confirmed writes consume a per-hour rate limit
(default 30/h). Anything above 20 items in batch_modify additionally reports
requiresConfirmation in the preview.

Engine performance

The C++ engine powers PapersGPT — the same index and
retrieval stack ships in this MCP server. Benchmark on a real Zotero installation
(full write-up):

• Indexing cost scales roughly linearly with library size; retrieval latency stays
  constant — a 10,000-paper library (30 GB) indexes in about 15–20 minutes, and
  everyday search stays at **15 ms**.
• For comparison: a typical web page load takes 1,000–3,000 ms; a blink of an eye is
  100–150 ms. PapersGPT answers in ~15 ms, fully offline.
• Privacy: your library never leaves your machine.

Configuration

Config lives at ~/.docsagent/config.json (or $DOCSAGENT_CONFIG) — one file shared by
the JS shell, the Python wrapper, and the C++ core. Validated against
spec/config.json.

Distribution

Both shells read the same config and talk to the same core — pick either (or both) as
your MCP distribution channel. Core lifecycle (start / stop / restart / status)
is available from both CLIs.

Links

• spec/ — the single-sourced contract: 8 tool schemas, 23 JSON-RPC methods,
  error codes + suggested calls, config schema
• PapersGPT search performance benchmark
• Zotero · Model Context Protocol

License

Apache-2.0