Local access with no logins — because your Zotero 7+ library is right there.
A local MCP server foryour Zotero desktop app.Free. Open source.
A native app that speaks MCP directly to Zotero 7 running on your machine. It lets desktop AI clients — whether they harness cloud models or run fully local models — read the library the Zotero Web API can't reach: your full PDF store without file-sync quotas, page-level figures and equations, live PDF annotations, the item you have selected right now, and Better BibTeX citation keys. Point it at a signed-in RefBridgeMCP account and Pro features — semantic search and the citation graph — light up on the same tool surface.
Download — v0.1.0
Latest public preview release. Builds are unsigned — on macOS right-click the app and choose Open the first time; on Windows click More info → Run anyway at the SmartScreen prompt. Signed builds are on the roadmap.
Requires Zotero 7 running locally with the built-in local API enabled (Settings → Advanced). See full release notes and checksums.
01. See the page, not just the text
Rasterizes any PDF page to an image on demand so multimodal models can read figures, tables, equations, and handwritten annotations. The Web API is text-only and can't do this.
02. Your whole PDF library, no quota
Streams full-text and attachment bytes from the local Zotero store. No Zotero file-sync subscription needed — a 50 GB library works the same as a 500 MB one.
03. "Summarize what I'm looking at"
Reads the item you have selected in the Zotero UI right now, plus every highlight and note on its PDF. Every AI-driven edit is snapshotted and tagged `ai-review` in Zotero, so you can one-click undo any change.
04. Better BibTeX aware
If you use Better BibTeX, tools return your real citation keys — the same ones you cite with in LaTeX, Pandoc, Obsidian, or Quarto — so AI-drafted prose is paste-ready.
05. No rate limits, no round-trip
Talks to 127.0.0.1 instead of api.zotero.org. No Web API throttling on large scans, and it keeps working on a plane, in a SCIF, or behind a locked-down campus network.
06. Yours to audit and fork
MIT-licensed, no account, no telemetry, no server in the middle. Read the source, pin a build, or self-compile. The only network traffic is your AI client calling its own model.
Which MCP clients can reach it
Any MCP client that runs on your machine can connect directly to the local endpoint the app prints on launch — no OAuth, no tokens, no tunnel:
- Desktop apps: Claude Desktop, ChatGPT desktop (developer mode), Cursor, VS Code (Copilot / Continue / Cline), Zed, Open WebUI.
- CLIs: Claude Code, Codex CLI, Gemini CLI, and anything else that speaks stdio or local HTTP MCP.
Browser-based chat like chatgpt.com or claude.ai runs in the cloud, so it can't reach 127.0.0.1 on your machine — for those, use the hosted version at refbridgemcp.com (or expose the desktop endpoint over a tunnel like Cloudflare Tunnel or ngrok, at your own risk).
Source & license
RefBridgeMCP Desktop is open source. Read the code, file issues, or fork it — the repository is live on GitHub.
License: GPL-3.0-only — use it, modify it, ship it, share your changes under the same license.