yilewang/llm-for-zotero

llm-for-zotero brings Large Language Models into the Zotero reader, so you can ask questions, summarize papers, inspect figures, compare sources, and save notes without leaving your library. It works with standard API providers, local OpenAI-compatible models, WebChat, Codex App-Server, and Claude Code.

Documentation:

• English
• Chinese

How to do a library chat literature review with llm-for-zotero agent

• At a Glance
• Quick Start
• What's New
• Configuration
• Demos
• File-Based Notes
• Agent Mode
• Skills
• General Web Search
• WebChat Setup
• Codex Setup
• Claude Code Setup
• MinerU PDF Parsing
• Privacy and Data Flow
• Roadmap
• FAQ
• Contributing

• Chat with the current PDF, selected text, figures, screenshots, and uploaded documents directly inside Zotero.
• Get grounded answers with citations that jump back to the source passage.
• Compare multiple open papers or add external files as extra context.
• Save answers, full conversations, and research notes to Zotero notes or local Markdown folders such as Obsidian and Logseq.
• Enable Agent Mode for library-wide read, search, tagging, metadata, import, note-editing, and organization workflows.
• Search the current public web and read relevant pages with Tavily, with source links attached to the answer.
• Use your preferred backend: API keys, local models, ChatGPT WebChat, Codex App Server, or Claude Code.

• Codex App Server is the recommended Codex path for ChatGPT Plus users. It runs through the local codex app-server runtime and is configured from the Agent tab.
• Claude Code Mode runs Claude Code as a separate conversation system inside Zotero through a companion local bridge. It is experimental and does not yet support native Zotero API operations.
• Skills let you customize how Agent Mode handles research workflows. The plugin ships with 8 built-in skills and a portal for creating your own.
• General Web Search lets the in-plugin Agent search the current public web with Tavily, read relevant pages, and attach source links to its answers.
• Standalone Window Mode opens the assistant in a dedicated window with paper chat, library chat, and conversation history.
• File-Based Notes save Markdown notes to local folders, including Obsidian, Logseq, or any plain Markdown directory.
• Cache-aware Agent Mode preserves stable paper context, prior read evidence, and coverage state across longer research turns, then compacts old transcript history automatically when the context window gets crowded.
• Citation navigation now keeps citation labels conservative until page locations are verified, while quote-based citations can jump back to the matching Zotero passage.
• MinerU PDF parsing provides higher-fidelity extraction for tables, equations, figures, and local mineru-api servers, with a richer file manager for bulk parsing, cache repair, sync packages, tags, and parsing filters.

Thanks to @jianghao-zhang and @boltma for major contributions to the Codex App Server, Claude Code, and file upload workflows.

1. Download the latest .xpi file from the Releases page.
2. In Zotero, open Tools -> Add-ons -> gear icon -> Install Add-on From File, then select the .xpi.
3. Restart Zotero.
4. Open Preferences -> llm-for-zotero, choose a provider, enter the base URL, key, and model, then click Test Connection.
5. Open a PDF in Zotero and click the LLM Assistant icon in the right-hand toolbar.

If you do not want to use a provider API key, start with WebChat or Codex App Server.

Open Preferences -> llm-for-zotero.

1. Select your Provider.
2. Paste your API Base URL, secret key, and model name.
3. Click Test Connection.

The plugin supports multiple provider protocols, including responses_api, openai_chat_compat, anthropic_messages, and gemini_native.

You can configure multiple providers and models for different tasks, such as a multimodal model for figures and a text model for summaries. The conversation panel also supports model-specific reasoning levels and hyperparameters such as temperature and max_tokens_output.

On the first message, the model loads the current paper as context. Follow-up questions use focused retrieval from the same paper, keeping conversations fast and grounded.

Click any generated citation to jump straight to the source passage in Zotero.

Summarize a full paper, focus on methodology or results, or select any paragraph and ask the model to explain it.

The selected-text pop-up can add highlighted text to chat with one click. It can also be disabled in settings.

Take screenshots of figures, attach up to 10 screenshots, or upload local files as additional context. Supported uploads include PDF, DOCX, PPTX, TXT, and Markdown.

Open multiple papers in Zotero tabs and type / to cite another paper as additional context.

Save answers or selected text to Zotero notes, export full conversations in Markdown, and customize quick-action presets for repeated research tasks.

Beyond Zotero's built-in notes, the agent can save Markdown research notes to any local directory you choose. Point it at an Obsidian vault, a Logseq graph, or a plain folder of .md files.

Open Preferences -> llm-for-zotero and scroll to the Notes Directory section.

Ask the agent to write a note using the configured nickname, for example: "Summarize this paper and save it to Obsidian." The agent gathers paper metadata, writes a Markdown note, adds YAML frontmatter, optionally copies figures from MinerU-parsed PDFs, and saves the note under the configured folder.

Or if you want to keep notes inside Zotero, the agent can also write to internal item notes with the write-note skill. Just ask it to "save a note for this paper" without mentioning an external directory.

Notes use Pandoc citation syntax such as [@citekey], which works with Obsidian Zotero Integration, Pandoc plugins, and many Markdown readers.

Note templates and figure-embedding rules live in the write-note skill. Open the Standalone Window -> Skills portal to edit them.

Agent Mode is disabled by default. Enable it in Preferences, then toggle Agent (beta) in the context bar.

It can read and search your library, draft notes, update metadata or tags with confirmation, and undo recent write actions from the conversation.

When enabled, the LLM can act on your Zotero library with read tools, write tools, confirmation cards, and undo.

Long agent runs are cache-aware. The plugin keeps stable Zotero context and previously read evidence separate from the changing chat transcript, tracks which papers and passages have already been inspected, and automatically compacts old turns when the model context fills up. This lets follow-up questions reuse grounded evidence when it is still relevant, while still asking the agent to read again when the needed source or coverage layer is missing.

The design philosophy is simple: read tools are unrestricted; write tools stay reviewable and undoable.

Use /discover_related while viewing a paper to run the direct OpenAlex action. It opens Recommendations, References, and Citations tabs with up to 20 papers per mode by default. Use /discover_related 30 to change the initial retrieval limit, Load more to retrieve 20 more per mode while preserving your selections, and Import selected to save the papers you choose. This shortcut does not require a model.

Ask “find relevant papers” in chat to start model-assisted discovery. The agent assesses retrieved candidates and presents a ranked card with relevance reasons. An explicit count, such as “find 15 relevant papers,” sets the batch size; otherwise it starts with five. Use Find more to add another batch of the same size while preserving your selections, then Import selected to save the papers you choose. Requests for references, citing papers, or a specific search source constrain discovery to that request. If fewer relevant matches are available, the card explains the shortfall instead of padding the list. A configured model is required for model-assisted discovery; literature search used to answer a question remains a separate flow that can finish with a cited answer without an import card.

Skills customize Agent Mode behavior for recurring research workflows such as paper QA, evidence retrieval, figure analysis, paper comparison, literature reviews, note writing, and cited-reference import.

Agent Mode can search the current public web with Tavily or AnySearch and read returned pages when search-result snippets are not enough. Use it for current facts, official documentation, news, finance, product information, and other general-web evidence. Scholarly discovery remains separate: the agent searches OpenAlex, arXiv, and Europe PMC (with CrossRef and Semantic Scholar for metadata lookup), and it can combine both kinds of search when a question needs academic and general-web sources.

To enable general web search with Tavily (the default):

1. Get a Tavily API key from app.tavily.com.
2. Open Preferences -> llm-for-zotero -> Agent.
3. In Web search, select Tavily, paste the API key and click Test connection.
4. Enable Agent Mode and ask it to search or verify something online.

There is no separate enable switch. The web_search and web_read tools become available to compatible in-plugin Agent conversations when Tavily has a key or AnySearch is selected. They are not added to WebChat, Codex App Server, or Claude Code conversations, which use their own runtimes and tool sets.

With Tavily, the agent chooses basic or advanced search and page-reading depth from the request, or follows an explicit request such as "use advanced web search." It can narrow searches by topic, date, or domain, and answers include clickable source indicators for the pages actually used.

Basic search costs 1 Tavily credit and advanced search costs 2. Page extraction also consumes Tavily credits, and Tavily currently offers a free monthly allowance. The API key stays in local Zotero preferences, but search queries and requested URLs are sent to Tavily, so do not include credentials or sensitive private text in web queries.

In Preferences → llm-for-zotero → Agent → Web search, select AnySearch. Leave AnySearch API key (optional) blank for anonymous access, or enter an existing key. Saving preferences does not make a request. The provider selection is persisted; selecting AnySearch does not reuse a Tavily key. No environment-variable key fallback is used. Test search (sends a public query) explicitly sends a fixed public query without invoking a model and displays result count, titles, URLs and the request ID when available. It consumes service quota; it is not an Agent conversation or proof of the Agent tool/citation path.

AnySearch supports general search with query and maxResults (1–10), and web_read with one to five URLs returned by search in the same run. Extraction uses at most two concurrent requests and returns bounded page text, not query-focused passages. Depth, topic, date/domain filters, chunk controls and vertical/source selection are not supported and are rejected rather than silently ignored. PDF, Office and media-binary extraction are not supported. Search queries and requested URLs are sent to AnySearch; do not include credentials or private library content.

Anonymous quotas and permissions are server-controlled. An exhausted quota produces an error; the integration does not adopt credentials from responses, automatically retry, switch providers, or report fictional credit usage.

If you have a ChatGPT Plus subscription, you can use Codex models in the plugin without a separate API key by signing in through the Codex CLI.

New users should choose Codex App Server from the Agent tab. The older Codex Auth (Legacy) path remains available for existing users, but is planned for future deprecation after app-server validation.

Claude Code mode runs Claude Code as a separate conversation system inside Zotero. It reuses the sidebar and standalone-window UI, but has separate conversation history, scope state, model settings, permission semantics, slash commands, and project skills.

Claude Code mode currently does not support native Zotero API operations. Use built-in Agent Mode for native library tools such as reading item state, editing notes, tagging papers, updating metadata, or importing items.

claude

git clone https://github.com/jianghao-zhang/cc-llm4zotero-adapter.git
cd cc-llm4zotero-adapter
npm install
npm run build
npm run serve:bridge

curl -fsS http://127.0.0.1:19787/healthz

./scripts/install-macos-daemon.sh

npm run daemon:status
npm run daemon:start
npm run daemon:stop
npm run daemon:restart
npm run daemon:uninstall

~/Zotero/agent-runtime/profile-.../

CLAUDE.md
.claude/settings.json
.claude/skills/
.claude/commands/

MinerU is an advanced PDF parsing engine that extracts high-fidelity Markdown from PDFs, preserving tables, equations, figures, and complex layouts that standard text extraction often mangles.

When enabled, the plugin sends newly added PDF attachments to MinerU for parsing and caches the result locally. Later interactions with that paper use the MinerU-parsed content.

1. Open Preferences -> llm-for-zotero.
2. Find the MinerU section and check Enable MinerU.
3. Keep cloud mode enabled, or check Use local MinerU server for local mode.
4. For cloud mode, optionally enter your own MinerU API key — see below.
5. For local mode, run a self-hosted mineru-api server and keep the default base URL (http://127.0.0.1:8000) unless your server uses a different address.
6. Add or import a PDF into your Zotero library. The plugin will automatically parse newly added PDF attachments with MinerU and cache the result for future conversations.

MinerU can start without an API key through the built-in API, but a personal key is strongly recommended. The built-in API may no longer be supported after June 1, 2026.

To get a free personal key:

1. Go to mineru.net and create an account.
2. Navigate to account settings and generate an API key.
3. In Zotero, paste the key into the MinerU section.
4. Click Test Connection.

When a personal key is provided, the plugin calls https://mineru.net/api/v4 directly.

Local MinerU server support was contributed by @renyong18 in PR #152.

Local mode sends PDFs to a self-hosted mineru-api server through POST /file_parse and stores the returned ZIP output in the same local cache format as cloud parsing. The default base URL is http://127.0.0.1:8000.

Prerequisites for local mode:

1. Install MinerU and run mineru-api (see the MinerU docs for installation).
2. Make sure required models are downloaded — mineru-api lazy-loads on first request, so the very first parse (or the first parse after switching backend) can take noticeably longer than steady state.

You can pick a Backend in the local section:

• pipeline (default) — general-purpose, multi-language, CPU-friendly.
• vlm — VLM-based, high accuracy on Chinese/English documents, requires GPU.
• hybrid — newer high-accuracy hybrid pipeline, multi-language, requires local compute.

The first parse after starting the local server, or after changing backend, can be slow while MinerU loads or downloads models. Test Connection checks that the server process responds at /health; it does not guarantee that all models are warmed up.

With the default 127.0.0.1 address, PDFs stay on your machine. If you change the base URL to a LAN or remote server, PDFs are sent to that server.

Pause / cancel limitation: mineru-api exposes no cancel or DELETE endpoint (only POST /file_parse, POST /tasks, GET /tasks/{id}, GET /tasks/{id}/result, GET /health). When you click Pause, the plugin stops the queue and aborts the HTTP wait, but the parse already running on the server keeps executing until it finishes — the GPU/CPU will not free up sooner. If you need to abort immediately (for example to switch backend without waiting), restart the mineru-api process yourself.

The MinerU preferences tab includes a Manage Files panel for maintaining parsed PDF caches:

• Browse cached and uncached PDFs by collection, tag, title, author, year, and added date.
• Start parsing all visible files, only filtered files, or selected files.
• Repair local MinerU caches and synced packages when metadata or files drift.
• Delete all, filtered, selected, or single-item caches from the manager.
• Use tag filters, including automatic Zotero tags, to choose which papers are included in bulk actions.

Advanced parsing filters can skip files before automatic or bulk parsing:

• Skip files over N pages controls the maximum page count used by Start All, Start Filtered, Start Selected, and auto-parse. The default is 100 pages.
• Exclude PDFs by Filename accepts comma-separated substrings, or regex patterns wrapped in /slashes/, for translated copies, supplements, or other files you do not want parsed automatically.

If Sync MinerU cache with Zotero file sync is enabled, the plugin can create companion ZIP attachments containing full.md, manifest.json, content_list.json, and extracted assets. Existing local caches sync only when you request it from the MinerU tab, and synced packages can restore a missing local cache when needed.

WebChat mode sends questions to chatgpt.com, chat.deepseek.com, and gemini.google.com through a browser extension, then streams responses back into Zotero. It is useful when you want browser chat access without a provider API key.

Prerequisites:

• Access to the selected provider site. Provider account requirements depend on the site and feature; Gemini permits some anonymous text conversations, while PDF upload and history may require sign-in.
• A Chromium-based browser such as Chrome.

Setup:

1. Download the latest extension.zip from sync-for-zotero releases.
2. Unzip it.
3. Open chrome://extensions, enable Developer Mode, choose Load unpacked, and select the unzipped extension folder.
4. In Zotero, open Preferences -> llm-for-zotero and set Auth Mode -> WebChat.
5. ⚠️: Keep the selected provider tab open in your browser. A green dot in Zotero means the extension and that exact site are connected. Make sure the tab and Zotero stay in the same monitor. No minimization or backgrounding, or the connection may drop.

Gemini requires a companion extension version that explicitly advertises the gemini target and DOM answer capture; Zotero rejects older extension builds before dispatch. The canonical gemini.google.com/app/:id conversation path, anonymous text flow, and signed-in PDF composer DOM have been observed in Chrome. A complete automated Gemini PDF, follow-up, history, and restore run remains required before a release is declared verified.

For release validation, keep Chrome signed in with the development extension loaded and run:

npm run test:webchat:live

This opt-in gate creates a real Zotero PDF fixture and clicks the real Zotero composer send control. It verifies one exact PDF upload and answer, toggles the visible PDF chip both ways, then verifies one prompt-only turn with zero submitted PDFs through the relay and Chrome extension.

Data flow depends on the backend you choose. Local models and local MinerU can keep processing on your machine; cloud providers, WebChat, Claude Code, Codex, and cloud MinerU involve their respective services or companion runtimes.

• Agent mode (beta)
• MinerU PDF parsing
• GitHub Copilot auth
• WebChat mode (ChatGPT, DeepSeek, and Gemini web sync)
• Standalone window mode
• File-based notes (Obsidian, Logseq, any Markdown directory)
• Claude Code integration
• Codex App Server integration
• Local MinerU support
• Customized skills
• Cross-device synchronization (MinerU cache)
• Agent memory system

Q: Does it require an API key to use this plugin?

A: It depends on the backend you choose. The plugin supports multiple backends with different requirements:

Q: Is it free to use?

Yes, the plugin is free. You only pay for API calls if you choose a paid provider. With Codex App Server, ChatGPT Plus subscribers can use Codex models without a separate API key. If you find this helpful, consider leaving a star on GitHub or buying me a coffee.

Q: Is my data used to train models?

The plugin does not train models. Data handling depends on the backend you choose: your configured API provider, local model, WebChat, Codex, Claude Code, or MinerU.

Q: How do I report a bug or ask a question?

Please open an issue on GitHub.

The plugin keeps routine diagnostics quiet by default while retaining warnings and errors. To collect more detail, open Zotero's Advanced Configuration Editor and change the string preference extensions.zotero.llmforzotero.logLevel.

Changes take effect immediately; an invalid value falls back to warn. Use Zotero's Help → Debug Output Logging controls to record or view the output, reproduce the problem, and then restore warn. This preference controls the plugin's diagnostic verbosity; Zotero's controls determine whether debug output is recorded or displayed. Other plugins' logging is unaffected.

Contributions are welcome. Bug reports, feature requests, documentation improvements, and pull requests are all useful. Please open an issue or submit a PR.

Run LLM_FOR_ZOTERO_TEST_ENTRIES=test-isolated/maintenance npm run test:workflow with the repository's native Zotero test environment configured. This suite measures idle maintenance first in a fresh disposable scaffold profile, then checks live diagnostic levels against Zotero's debug output. Keep it separate from deletion workflows, which intentionally leave pending cleanup and retry work.

Model context limits and provider-defined reasoning options are maintained in registry/model-capabilities.v1.json. The plugin refreshes this schema-validated registry and configured provider catalogs in the background, with a bounded first-use refresh when needed.

Auto — provider default leaves reasoning controls to the endpoint. An unfamiliar model offers Auto until capability information is available; missing metadata never means reasoning is off. Explicit levels, including Off, come from the applicable profile or endpoint capabilities. Saved choices are checked again when the model or endpoint changes, and unavailable choices resolve to Auto. Explicit advanced request-body parameters retain their existing precedence.

User profile overrides take priority over endpoint metadata, followed by the remote/bundled registry and established model profiles. A catalog's supports_reasoning boolean does not establish an effort list. The standard OpenAI /models API does not publish supported efforts. Custom endpoints may opt into the plugin's structured catalog extension: a model row's reasoning object uses the registry's validated kind, options, and optional defaultOptionId contract. An option can declare a portable effort string for OpenAI-compatible protocols or an existing declarative controls patch for its endpoint. Capability data is cached per endpoint, protocol, authentication mode, and runtime scope.

Adding a verified model or effort to the remote registry does not require another plugin release once the client supports its encoding. Increment the registry revision, run npm run validate:model-registry, and publish the JSON change through the normal reviewed update process. This does not infer undocumented future API contracts or automatically publish model settings.

Standalone MCP clients can use the existing bearer-authenticated /llm-for-zotero/mcp endpoint while Zotero is running. Authenticated clients can use the exposed library, file, and command tools without additional Zotero permission switches or directory grants. File and command access uses the operating-system permissions of the Zotero process. The connected assistant owns approval through its own permission settings; the plugin does not apply Original Agent Safe/Auto/YOLO or display a second permission prompt. Integrated Codex and Claude Code use their existing MCP enablement controls and the same delegated approval rule.

Use ordinary tools/call requests; no Zotero chat, private turn token, or additional session handshake is required for standalone clients. Specify libraryID for predictable targeting, or omit it to resolve the currently selected library once for that call. Invalid arguments, unavailable targets, native read-only restrictions, and execution or verification failures remain errors. Writes retain durable recovery records and native Zotero verification; preserve returned action IDs for recovery. Standalone clients recover through the undo tool: pass actionId to undo one action, or actionIds to revert several (which cannot be combined with count). Do not blindly repeat a write after a timeout or uncertain outcome: inspect native state and the returned recovery information first. This interface does not promise exactly-once execution across repeated HTTP requests.