Expert books for Claude Code, with sources cited.
Connect expert books and your own docs to Claude Code or ChatGPT. Your AI reads the right pages, then cites them.

How do you give Claude Code knowledge from expert books?
Put the books and docs you trust in a library, and connect that library to Claude Code, ChatGPT or any AI tool that speaks MCP. Then your AI opens the relevant pages before it answers, and cites them.
You can build this yourself with free tools such as Zotero with Zoteus, or pdf-mcp. Or you can use Trove: one library of expert books and your own docs, read by every AI you use. Connect Claude or ChatGPT in one click. Coding tools take one install command per machine. There is no server to host.
What goes in
- Expert books from the marketplace
- Your notes, guides and playbooks
- PDFs and files you already own
Who reads it
- Claude and Claude Code
- ChatGPT and Codex
- Cursor, Grok, GitHub Copilot
What does a cited answer look like?
Here is the shape of it, using a real book and a real page from the Trove marketplace. The agent reads the chapter, answers in its own words, and cites its source so you can check it.
Our support agent needs to answer from our product docs. Should we build RAG over them, or let the agent read them?
For open-ended questions, let the agent read. Chunked retrieval returns fragments, can't follow a “see chapter 12” reference, and goes stale when the docs change.
An index still makes sense for a narrow, stable set of repeat questions, like a fixed FAQ, where it's cheaper per query.
- Building Agents in Production · p. 3 · Tools Over RAG
Why isn't the model enough on its own?
Claude Code and ChatGPT learned from a lot of the web. They haven't read the best books on your topic, many of which aren't on the open web, and they haven't read your own notes, guides and decisions.
So on a hard question the model answers from general patterns. It sounds sure, and you can't tell which part came from a real source and which part it filled in.
The fix is not a bigger model. It's giving the model the right pages to read.
How do you do it yourself with Zotero, Zoteus or pdf-mcp?
If you already own the PDFs, free tools can connect them to your AI. As of November 2026:
- Zotero is a free, open-source reference manager from a nonprofit. It keeps your library, files and notes. On its own, it doesn't connect to AI assistants.
- Zoteus is an MCP server that connects Claude, ChatGPT, Cursor, Codex and other apps to your Zotero library, with page numbers in its answers. It's free to self-host (MIT licence), with hosted plans from €7 a month and lab and department plans above that. It's built for researchers and review teams.
- Zotero MCP is a free, open-source Zotero MCP server on GitHub.
- pdf-mcp is a free, open-source MCP server that runs on your machine and searches a folder of PDFs, with page-level results, for Claude Code, Cursor, Codex and other MCP clients.
The catch is everything around them. You find and buy the books yourself. You install, configure and update a server on each machine, and connect it in each AI tool. And your AI only ever knows what you've already gathered.
DIY tools vs Trove: what's the difference?
| Question | DIY (Zotero + Zoteus, pdf-mcp) | Trove |
|---|---|---|
| Where the books come from | PDFs you already own | Expert books from the marketplace, plus your own files |
| Your own notes and docs | As PDFs or Zotero notes | Uploaded, or written by your agent as a Manuscript |
| Setup | Install and update a server on each machine | Claude or ChatGPT in one click; coding tools one install command per machine; no server to host |
| AI tools | Any MCP client you configure | Claude, Claude Code, ChatGPT, Codex, Cursor, Grok, GitHub Copilot |
| Citations | Yes. Page-level results | Yes. Every answer cites its source |
| Share with teammates | Partly. Zoteus Lab plan shares a group library; pdf-mcp runs on one machine by default | Yes. Team Library |
| Cost | Free self-hosted; Zoteus hosted from €7 a month | Free to start; see pricing |
Which should you pick?
DIY can work if
- You already own every PDF you need
- You work on one machine
- You're happy to run and update a server
Trove fits if
- You want expert books you don't own yet
- You use more than one AI tool
- You want one click for Claude or ChatGPT, one install command per machine for coding tools, and no server to host
- You'll share books with teammates
How do you start with Trove?
Create a free account
Start on the free personal plan. No card needed.
Connect your AI tools
Follow the install guide for Claude, Claude Code, ChatGPT, Codex, Cursor, Grok or GitHub Copilot.
Add a book and ask a real question
Pick a book from the marketplace or upload your own, ask something you know the answer to, and check the page it cites.
Compare plans in the pricing section. Working with a team? Read about shared playbooks for every teammate's AI.
Questions, answered.
Can Claude Code read PDFs and books on its own?
It can read a file you point it at inside your project. It can't open books you don't have, and it doesn't keep a library across projects or share one with your other AI tools. A library connected through MCP fills that gap.
Does this work with ChatGPT too?
Yes. Zoteus lists ChatGPT among the apps it connects to, and Trove connects to ChatGPT as well as Claude, Claude Code, Codex, Cursor, Grok and GitHub Copilot. pdf-mcp runs locally, so it suits local tools like Claude Code, Cursor and Codex.
Will the answers cite their source?
With Trove, every answer cites its source, so you can check it. When two sources disagree, you see both. Zoteus and pdf-mcp also return page-level results from your own PDFs.
Is my content used to train AI models?
Not with Trove. Your library is private by default, we never train models on your content, and when you delete a book it's gone. With a self-hosted DIY setup, your files stay on your own machine.
What does Trove cost?
It's free to start, with personal and team plans. See the pricing section on the homepage for current plans.
Turn your AI into an expert
Give your AI the books it hasn't read.
Start free, connect the tools you already use, and check every answer against its source.