Zotero AI Butler: LLM paper summaries as native Zotero notes

The backlog problem this plugin is aimed at

The README opens with a line that will be familiar to anyone with a large Zotero library: papers saved for later reading that never get read. It names three specific failures. Too many articles to read, and sending them to a chat model one at a time is slow. Forgetting the contents of a paper two days after finishing it. Losing the thread in long papers even with a translation plugin available. The plugin's answer is to move the summarisation step into the library itself, so that the note lands on the Zotero item rather than in a separate chat window. The stated audience is researchers and students who already use Zotero as their reference manager and want the reading step partly automated. The README is explicit that this is a third-party plugin that does not proxy any model traffic: you apply for and configure your own API key, and requests go from your machine to the provider you chose.

Producer, consumer, and where the PDF actually goes

The README describes analysis as running on a producer-consumer pattern, with all tasks visible in a queue page inside the plugin dashboard. Tasks can be pending, in progress, completed or failed. Three things can put a paper into that queue: a right-click menu item on a single item, an automatic scan of newly added PDFs, and a bulk scan of items that have no AI Butler note yet, which the dashboard arranges by your existing Zotero collection structure. The automatic scan is off by default, and the README says this is to minimise impact on Zotero's performance. When enabled, the plugin waits for Zotero to finish retrieving metadata before it starts. That ordering matters, because the prompt templates reference variables such as title and authors, which come from the item metadata rather than the PDF. On the model side there are two PDF handling modes. Multimodal mode uploads the PDF as Base64, which the README says lets the model see the original layout, including figures, formulas and tables. Text extraction mode is for models that cannot accept images. The README claims the multimodal path also covers PDFs that are mostly images, which is the case where text extraction would return little or nothing.

Installing it and configuring a provider

Installation follows the normal Zotero plugin route from the releases page, and the README points to a separate wiki for configuration and usage documentation. Once installed, the dashboard opens either from Zotero's Edit menu, under Settings and then the AI Butler tab, or from a right-click on any item via the AI Butler dashboard entry. The settings page is where the work happens. Under API configuration the README lists OpenAI, Gemini, Anthropic and an OpenAI-compatible option for third-party providers, with Volcengine Ark and Ollama also appearing in the platform table. After entering a key there is a test connection button to verify it works. Two other keys sit on the same page. Task processing speed sets how many papers are handled per minute, which the README frames as a way to avoid hitting API rate limits. PDF processing method selects between multimodal Base64 and text extraction. Prompt templates are also configurable, with built-in presets, custom presets, a live preview of variable substitution, and a restore-to-default action. The README does not give the exact JSON or preference keys behind these controls, so if you need to script or audit the configuration you will have to read the wiki or the source.

What the notes contain, and what the multi-round mode changes

The default output is a Markdown note saved under the Zotero item, and the README notes that interactions with it behave like a native Zotero note, so it can be opened alongside the PDF. Beyond the single-pass summary there is a deeper reading mode that walks through background, methods, experiments and conclusions in sequence. The multi-round mode is the more interesting design choice: instead of one prompt, you define a prompt for each round, covering research background and question, methods and techniques, experimental design and results, and conclusions and outlook, and the plugin then merges the rounds into one final note. A right-click option called multi-round re-reading offers two strategies, described in the README as multi-round concatenation and multi-round summarisation, and both overwrite any existing AI note on that item. That overwrite behaviour is worth knowing before you run it on a paper whose note you have already edited by hand. The README does not state whether a backup of the previous note is kept.

The features that are not just summarisation

Three additions go past text summaries. A mind map is generated for each paper and can be zoomed, exported as PNG, or exported as OPML for use in other outliners. A multi-paper review mode runs from the right-click menu on a collection, produces a literature review report, and creates a separate report item linked to the original PDFs. A one-image summary mode generates an academic poster style image per paper using Nano Banana Pro, which the README presents as a way to build a visual memory anchor for the paper. There is also a sidebar for reading, with LaTeX rendering and follow-up questions against the summary, and the README mentions pinning the sidebar so it persists when you switch items. The follow-up conversation feature is labelled pre-release in the README, so treat it as less settled than note generation.

Where the design puts the burden on you

The plugin deliberately does not supply model access. That keeps the project free of hosting costs and keeps your papers off the maintainer's servers, but it means the setup cost is a provider account, a key, and whatever that provider charges. The README's own recommendation is Gemini 3 Pro for summarisation, and it links to a discussion post about deploying a proxy to obtain free Gemini access. That is a user-contributed workaround, not something the plugin ships, and it adds a component you have to run and keep current. The second burden is prompt quality. Since every template is editable, the note you get reflects the template you leave in place, and the README offers no evaluation of which preset performs best. The third is the release channel: the most recent releases listed are v4.1.0-beta.1 and v4.1.0-beta.2 from July 2026, with the last push to the repository in September 2026. If you want a version without beta in the name, you are choosing an older release. The README also does not describe what happens when a task fails beyond marking it failed in the queue, so retry behaviour on a bad key or a rate limit is something to check in the interface yourself.

Compared with doing it in a chat window or a script

The obvious alternative is exporting a PDF and pasting it into a chat interface, or writing a small script against a model API. The difference is where state lives. A chat window has no link back to the citation, so the summary drifts away from the library, and you re-upload the same paper every time you want to revisit it. A script can automate the calls, but you then own the queue, the retry logic, the rate limiting and the note formatting. AI Butler's contribution is that the queue, the per-minute throttle, the metadata variables and the note storage are already wired into Zotero, and the note is attached to the item it describes. The trade-off is that you inherit the plugin's choices: its supported providers, its two PDF handling modes, its note format. If you need a pipeline that also feeds a reference manager other than Zotero, or that writes into a knowledge base outside Zotero, a script gives you that and this plugin does not.

Licence, maintenance and what to check before trusting it with a library

The project is licensed AGPL-3.0. For a Zotero plugin that you install and run locally, that is mostly a question of what happens if you redistribute a modified build; the licence requires source availability for modified versions, and I am not giving legal advice on how that applies to your situation. The practical maintenance question is different. The plugin depends on external model APIs whose names and request formats change, and the README's platform table already lists model identifiers that will age. The repository shows a beta-heavy release cadence, so expect to update the plugin when a provider changes an endpoint. On data handling, the README's privacy statement is specific: no collection, storage or upload of your data, papers or API key by the plugin, with requests going directly from your device to your provider. That claim is about the plugin, not about the provider, and the provider's own retention policy applies to whatever you send. If you choose multimodal mode, what you send is the PDF itself, not extracted text. That distinction should decide your mode choice more than any quality claim in the README.

Editorial conclusion

Adopt it if you already keep a Zotero library and are willing to hold your own API key for Gemini, OpenAI, Anthropic, an OpenAI-compatible endpoint, Volcengine Ark or a local Ollama instance, and if you accept that the plugin's output is only as good as the prompt template you leave in place. Do not adopt it if you cannot send paper text or full PDF bytes to a third-party model, or if you need a stable release rather than the 4.1.0 beta line. Before installing, verify three things: that your Zotero version is supported by the current release, that the connection test passes for your chosen platform, and that the text extraction mode you pick actually handles your PDFs, since scanned or figure-only papers are the case most likely to fall over.

Official sources

1. License: AGPL-3.0
2. Project website
3. README
4. Releases
5. steven-jianhao-li/zotero-AI-Butler on GitHub