Upgrade StorageYet another Zotero AI plugin — BibGenie, built around checkable citations and local-first
Hi, yes, I know — the last thing the world needs is another Zotero AI plugin. I'm Bai, one of the people building BibGenie (bibgenie.com, formerly Zotero Copilot), and we've been working on it for over a year now. I figured I'd post here anyway, because we tried to do two things differently, and I'd rather hear from actual Zotero users whether those two things matter.
The pain point is pretty specific. Like a lot of people here, I was asking AI tools about papers and getting answers that sounded perfectly confident — and then having no idea which paper, let alone which page, the conclusion actually came from. So I'd end up flipping through the PDF myself to check, which kind of defeated the purpose.
So BibGenie is a native Zotero 7 plugin, sidebar-style, that works directly inside your library. The two things we spent the most time on are:
1. Citations you can actually trace back to the text block. This is the one I care about most. When BibGenie cites something, you can click it and jump straight to that page in the PDF, with the exact cited text block highlighted. EPUBs and snapshots resolve to the paragraph the same way. In a research workflow, I think that's the difference between an answer you can use and an answer you still have to verify by hand.
2. Local-first, with BYOK. Chat history lives in your local Zotero database, not on our server, and the library's semantic index is built locally. For models, you can pick any of three routes: bring your own key (OpenAI, Claude, DeepSeek, OpenRouter, and other compatible APIs — requests go directly to the provider and don't consume credits), run local models (Ollama / LM Studio, with both chat and embeddings local, so nothing leaves your machine), or just use the official model out of the box with zero configuration.
Beyond those two core principles, a few other things we built for daily workflow:
Granular context without file uploads: Use @ or right-click Add to Chat to pull in specific items, selected PDF pages, EPUB chapters, annotations, notes, tags, or a whole collection, and compare multiple papers side by side.
Zero-config built-in MCP server: Runs locally at 127.0.0.1:23119/bibgenie/mcp with no separate Python service needed. Claude Code, Claude Desktop, Cursor, and Cherry Studio can search your library, read PDFs (including rendered pages for charts and scanned pages), and read notes and snapshots.
Library cleanup with diff previews: Clean up collections, tags, and metadata checked against arXiv, Semantic Scholar, OpenAlex, DBLP, and OpenReview — every change comes as a diff preview first, and nothing is written to your database until you approve it.
On pricing, I'll be candid because this is a small indie product and there's no point pretending otherwise: all features are free, and if you use your own API keys or local models, it costs nothing. The only thing you pay for is if you use our hosted model tokens. Free users get 200 credits a month, and Pro is $9.99/month with 10,000 credits (more than enough for most people). There's also a $199 Lifetime Pro tier for early supporters, which will be retired for new purchases after v1.0 (existing buyers won't be affected at all). Because the plan includes recurring model credits, a one-time payment is hard for a small product to sustain once we move past the early-adopter stage.
For the official hosted models, requests are routed via OpenRouter, and we only add 7% on top of raw cost to cover server and payment processing expenses. We're not planning to make money off the token spread — especially since custom keys and local models are fully supported anyway.
Docs are at https://www.bibgenie.com/docs. I'd really appreciate feedback, especially from people who actually live in Zotero every day. Criticism is completely welcome — a lot of this product was shaped by users telling us what was wrong.
The pain point is pretty specific. Like a lot of people here, I was asking AI tools about papers and getting answers that sounded perfectly confident — and then having no idea which paper, let alone which page, the conclusion actually came from. So I'd end up flipping through the PDF myself to check, which kind of defeated the purpose.
So BibGenie is a native Zotero 7 plugin, sidebar-style, that works directly inside your library. The two things we spent the most time on are:
1. Citations you can actually trace back to the text block. This is the one I care about most. When BibGenie cites something, you can click it and jump straight to that page in the PDF, with the exact cited text block highlighted. EPUBs and snapshots resolve to the paragraph the same way. In a research workflow, I think that's the difference between an answer you can use and an answer you still have to verify by hand.
2. Local-first, with BYOK. Chat history lives in your local Zotero database, not on our server, and the library's semantic index is built locally. For models, you can pick any of three routes: bring your own key (OpenAI, Claude, DeepSeek, OpenRouter, and other compatible APIs — requests go directly to the provider and don't consume credits), run local models (Ollama / LM Studio, with both chat and embeddings local, so nothing leaves your machine), or just use the official model out of the box with zero configuration.
Beyond those two core principles, a few other things we built for daily workflow:
Granular context without file uploads: Use @ or right-click Add to Chat to pull in specific items, selected PDF pages, EPUB chapters, annotations, notes, tags, or a whole collection, and compare multiple papers side by side.
Zero-config built-in MCP server: Runs locally at 127.0.0.1:23119/bibgenie/mcp with no separate Python service needed. Claude Code, Claude Desktop, Cursor, and Cherry Studio can search your library, read PDFs (including rendered pages for charts and scanned pages), and read notes and snapshots.
Library cleanup with diff previews: Clean up collections, tags, and metadata checked against arXiv, Semantic Scholar, OpenAlex, DBLP, and OpenReview — every change comes as a diff preview first, and nothing is written to your database until you approve it.
On pricing, I'll be candid because this is a small indie product and there's no point pretending otherwise: all features are free, and if you use your own API keys or local models, it costs nothing. The only thing you pay for is if you use our hosted model tokens. Free users get 200 credits a month, and Pro is $9.99/month with 10,000 credits (more than enough for most people). There's also a $199 Lifetime Pro tier for early supporters, which will be retired for new purchases after v1.0 (existing buyers won't be affected at all). Because the plan includes recurring model credits, a one-time payment is hard for a small product to sustain once we move past the early-adopter stage.
For the official hosted models, requests are routed via OpenRouter, and we only add 7% on top of raw cost to cover server and payment processing expenses. We're not planning to make money off the token spread — especially since custom keys and local models are fully supported anyway.
Docs are at https://www.bibgenie.com/docs. I'd really appreciate feedback, especially from people who actually live in Zotero every day. Criticism is completely welcome — a lot of this product was shaped by users telling us what was wrong.