AI4Paper: a Zotero 7-10 plugin and web workbench for translating, reading and reviewing papers

What AI4Paper actually solves, and for whom

The project targets a specific bottleneck: reading and organising English-language literature when your working language is Chinese. Its README frames the product as a closed loop of finding, reading, managing and reviewing papers, and the concrete pieces are a PDF full-text translator that preserves layout, formulas and figures, a search layer that expands Chinese keywords into several English query variants, and journal metadata columns showing CAS partitions, JCR quartiles and impact factors directly in the list view.

The intended user is a Zotero-heavy researcher. The plugin declares support for Zotero 7 through 10, and the README describes installing an XPI through Zotero's own add-on dialog rather than through a separate application. A second audience is people who do not want to install anything: the browser workbench at app.ai4paper.pro is described as requiring no plugin. A third surface, a WeChat mini program, handles journal subscriptions and new-paper alerts.

The scale claim is the 240 million document source plus a citation-network index, which the README says backs the search, citation expansion and gap analysis. That index is the vendor's own, not a public dataset you can inspect, which matters for anyone who needs to audit where a claim came from.

How the plugin, the index and the AI services fit together

Two mechanisms are visible from the documentation. The first is metadata enrichment: the plugin extracts a DOI from a PDF, queries for metadata, then writes back title, authors, abstract, DOI, citation count, journal partition and impact factor, and can translate an English title into Chinese. The README lists this under refresh metadata and refresh journal partition, with custom columns for display. The second is retrieval plus generation: a Chinese keyword is expanded into multiple English search strings, run across several academic databases, ranked by citation count, and the results feed either a one-click import into Zotero or a review generator that crawls the citation network, filters in several rounds and produces a 3000 to 5000 word report with literature numbers attached to each claim.

Model access is deliberately provider-agnostic. The README lists more than 20 services, including domestic Chinese options such as Tongyi Qianwen, DeepSeek, Zhipu, Kimi, Doubao and SiliconFlow, international ones such as OpenAI, Claude, Gemini, Mistral, xAI, OpenRouter, Groq and Together AI, and local Ollama for fully offline use. The vendor's own AI4Paper service is marked as recommended. That means the quality of translation, notes and reviews depends on which endpoint you point it at, not on a fixed model the project controls.

A detail worth noting for anyone who cares about citation integrity: the README describes a fake-reference checker that takes a pasted bibliography and verifies each entry against authoritative databases. That is a useful counterweight to AI-generated text, though it also implies the review output needs verification by default.

Installing the Zotero plugin and running a first translation

There are two entry paths. The web path needs no installation at all: open app.ai4paper.pro, register, and start reading. The plugin path is the one with steps, and the README gives them explicitly.

Download the XPI from the releases page, then install it from Zotero's add-on dialog:

# 1. Download the latest XPI from:
#    https://github.com/wdcpclover/ai4paper/releases
# 2. In Zotero: Tools -> Add-ons -> gear icon -> Install Add-on From File
# 3. Select the downloaded .xpi file, then restart Zotero

After restarting, the README says to register an account to get a 7-day full-feature trial, and that one account works across the web workbench, the plugin and the mini program. If you prefer the browser route, the whole first use is a single URL:

# Open the workbench, register, and start reading
https://app.ai4paper.pro

For a first real task, the lowest-friction test is full-text translation. The README describes it as zero-configuration: open a PDF, trigger the full-text translation, and the output keeps the original layout, figures and formulas with the translation alongside. If you want to confirm the metadata pipeline instead, import a folder of PDFs and run the one-click organise step, which extracts DOIs, queries metadata, renames and files the documents. Expect the journal partition columns to be empty on entries whose DOI lookup fails, since that data comes from the enrichment step rather than from the PDF itself.

Where AI4Paper is the wrong tool

The licence is the first constraint. The README's licence section reads "Copyright © 2026 AI4Paper. All rights reserved." There is no open-source licence file in the repository root, which contains only CHANGELOG.md, README.md and privacy.html. So despite being hosted on GitHub with public releases, this is a proprietary product. If your institution requires an OSI-approved licence, or if you need to fork, self-host or audit the code, AI4Paper does not qualify. The repository's primary language being HTML is consistent with a documentation and release shell rather than a source distribution.

Second, the heavy features are service-dependent. The 240 million document index, the citation network, the review generator and the default AI service all run on the vendor's infrastructure. The README offers Ollama for fully offline model inference, but that covers the language model only; it does not make the literature index local. Anyone working in an air-gapped environment should treat the search and review features as unavailable.

Third, the value proposition thins out for English-native readers. Full-text translation, Chinese keyword expansion and Chinese title translation are the distinctive parts. If you read English papers directly and already have a reference manager configured, you are paying the setup and account cost for features you will not trigger.

Finally, the README does not document rollback, data export guarantees or what happens to cloud-synced notes if you stop paying. Notes sync is described as end-to-end encrypted, which is a security claim, but the operational exit path is not described anywhere in the README.

How it differs from Zotero GPT, PapersGPT and ZotSeek

The comparison that matters is with the Zotero AI plugins people already search for. Zotero GPT, PapersGPT and ZotSeek are the names that surface alongside AI4Paper, and the architectural difference is where the retrieval happens.

Those tools generally attach a model to what is already in your library. You select items, and the model answers over those items. AI4Paper inverts the direction: it ships its own literature index and citation network, so search, citation expansion and gap analysis query an external corpus before anything reaches your library. That is why the review generator can claim to attach a real literature number to each statement, and it is also why the tool cannot work without the vendor's service.

The second difference is breadth of surface. AI4Paper is not only a plugin. The same account covers a browser workbench, a WeChat mini program for journal subscriptions and a browser extension that shows journal partitions on Google Scholar and CNKI pages. A single-purpose Zotero plugin does one thing and does it inside Zotero; AI4Paper is closer to a small suite with Zotero as one client. If you want a minimal tool that never leaves your library, that breadth is overhead rather than a feature.

Maintenance cadence, release cost and licence implications

The repository was last pushed on 2026-09-19, and the releases list shows plugin-v8.2.180, plugin-v8.2.179 and plugin-v8.2.178 all dated 2026-09-19. Three releases in a single day is a fast patch cadence, and the version numbering suggests frequent incremental builds rather than occasional milestones. The repository is not archived. For a plugin that must track Zotero's own API, that cadence is the relevant signal: Zotero 7 through 10 support is a moving target, and a plugin that stops shipping will break on the next Zotero release.

The upgrade cost for users is low by design. Installing a new XPI over the existing one is the documented path, and the account is shared across surfaces, so there is no per-device configuration to redo. The risk is on the other side: because the plugin is distributed as a binary XPI and the repository root holds no source, you cannot diff what changed between v8.2.179 and v8.2.180. You are trusting the release, and the CHANGELOG.md is the only public record of what moved.

On licensing, the all-rights-reserved notice means you have no redistribution rights and no source access. That is a commercial arrangement, not a legal problem, but it changes procurement: institutional review boards and university IT departments that require licence review will need the vendor's terms, and the README does not link to them. The privacy.html file in the repository root is the only policy document present. Whether your institution accepts a hosted service that stores your library metadata and notes is a decision the README does not help you make.

Editorial conclusion

Adopt AI4Paper if you already run Zotero 7 through 10 and want PDF translation, journal partition columns and AI notes inside the library you already have, or if you want the same account on a browser workbench with no plugin. Do not adopt it if you need an open-source, self-hostable tool, if your workflow is entirely English-language, or if you cannot accept a hosted service with a 7-day full-feature trial and no published source licence. Before committing, verify three things: that your Zotero version is in the 7-10 range, that the XPI you download is the newest release from the project's releases page, and that the vendor AI service or one of the listed providers is reachable from your network, since the plugin depends on an external model endpoint for most of its features.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Releases
5. wdcpclover/ai4paper on GitHub