NotebookLM vs Claude Projects vs MCP: studying with citations
LearnOS Team, October 14, 2026 · 6 min read
"Make the AI answer from my notes, and tell me exactly where it got that." It's the right thing to ask of a study assistant, and in 2026 there are three common ways to get it: upload your material into a notebook app, put it in a chat project, or connect your library to your AI tools over MCP (the Model Context Protocol), which is how LearnOS works.
They all retrieve from your material before answering. Where they differ is in how precise the citation is, where your material lives, and which AI you're allowed to ask. This post compares them on those three things. Facts are as of October 2026, with links to each vendor's docs.
One naming note first: Google renamed NotebookLM to Gemini Notebook on July 16, 2026 (Google). Same product, new name. We use the new name below.
The one-line version of each
- Gemini Notebook (formerly NotebookLM). A notebook app built around your sources. Chat answers are grounded in what you uploaded, and citations are direct quotes you can hover over or click to see in context. It's the best citation experience of the three.
- Claude Projects. Your files become "project knowledge" in Claude. On paid plans, large projects switch to retrieval automatically, so Claude searches for the relevant passages instead of loading everything. Great if you already work in Claude; Anthropic's docs don't promise passage-level citations.
- MCP. Your library stays in one place and is exposed as tools that any MCP client can call: Claude Code, Cursor, Claude itself through custom connectors, and others. The client's model writes the answer from what the tools return. Citation quality depends on what the server returns and on your client.
How they compare
| Gemini Notebook | Claude Projects | MCP (e.g. LearnOS) | |
|---|---|---|---|
| Where your material lives | In each notebook | In each project | One library, outside any one AI app |
| Which AI answers | Gemini, in Google's apps | Claude | Whatever MCP client you use |
| Citation form | Direct quotes; hover for the full quote, click to jump to it in context | Not documented as passage-level; ask Claude to cite and spot-check | Whatever the server returns (LearnOS: course, lesson, heading and a link to that section) |
| Grounding | Answers from your sources only | Retrieves from project knowledge | Retrieves through tool calls; the client decides how to use the results |
| Scale | Up to 500,000 words or 200 MB per source; capped sources per notebook | Up to ~10x more knowledge with retrieval, paid plans | Set by the server |
| Setup | Upload files | Upload files | Add one server URL to your client |
Each column describes a position, not a grade. Which one fits depends on how you study.
The differences that actually matter
How precise the citation is. Gemini Notebook leads here. Google's help center says it "uses direct quotes, text, and images from your sources as citations". You can hover a citation for the full quote, or select it "to navigate directly to the quote and review it in context" (Google). One caveat from Google's FAQ: if a source is very short, it cites the whole document rather than a passage (Google).
Claude Projects retrieves with a "project knowledge search tool" once a project outgrows the context window (Anthropic). That keeps answers grounded in your files, but the help article doesn't describe clickable, passage-level citations. In practice, tell Claude in the project instructions to name the file and section for every claim, then spot-check a few.
With MCP, the citation is only as good as what the server returns. A server that hands back whole files gives your AI nothing precise to cite. A server that returns heading-level sections, with a stable ID and URL for each one, lets any client cite the exact section. That's the design choice to look for. (If you're weighing retrieval against long context in general, see Is RAG still worth it with long-context models?)
Where your material lives. Notebooks and projects are containers inside one product. Studying across three courses means three notebooks or projects, or one giant one. Google's July announcement brings notebooks into the Gemini app and, later, AI Mode in Search; it doesn't mention third-party AI tools. Claude Projects are a Claude feature.
An MCP server inverts this. The library is the fixed point, and the AI is interchangeable. The same courses, notes and flashcards are available whether you're in Claude, in Cursor while you code, or in a different client next year.
Which AI you can use. This follows from the last point. Gemini Notebook answers with Gemini; Projects answer with Claude. MCP is an open protocol, so any client that supports it can connect. Anthropic's own custom connectors accept remote MCP servers on Free (one connector), Pro, Max, Team and Enterprise plans, as long as the server is reachable on the public internet (Anthropic).
What the setup costs you. Uploading files is the easiest path, and it's why notebook apps win so many "how do I…" answers. MCP is one more step: add a server URL to your client, and sign in once if it uses OAuth. Building your own MCP server over a folder of notes is possible too, but then chunking, search quality and citation IDs are your problem.
Doing it with MCP: LearnOS
LearnOS is a learning platform with a hosted MCP server. It's the MCP column above, with these specifics:
- Sign-in: add
https://learnosapp.com/mcpto your client and approve access in the browser (OAuth). No API keys to paste. - Citations: lessons are indexed as heading-level sections. Every search result carries the course, lesson, heading, a
lesson_idand an anchor, plus a URL to that exact section, so your AI can cite the section and you can open it. - Your model writes the answer. The server only retrieves; it never generates the answer. Search and reading cost nothing extra. Generating slide decks or audio overviews counts toward LearnOS's free usage limits.
- Honest limits: you won't get Gemini Notebook's hover-to-quote cards inside your AI client. You get links to sections. And it works with the courses in your LearnOS library, not arbitrary PDFs dropped into a chat.
So which should you use?
- One subject and a pile of PDFs, and you want the best citation UI: Gemini Notebook.
- You already live in Claude and want answers from your files in chat: Claude Projects (retrieval for large projects needs a paid plan).
- You want one course library you can reach from whichever AI tool you use, including your coding agent, plus notes, flashcards and progress that stay with it: an MCP server, LearnOS or your own.
Whichever you pick, test it the same way: ask a question whose answer you know, then open the citation and check that the passage actually says that.
Go deeper: learn how MCP works
If the MCP column is new to you, MCP for Beginners (Microsoft's open course, which we've ported to LearnOS) takes you from the core concepts to a deployed server:
- Learn the core concepts: hosts, clients, servers, tools, resources and prompts.
- Build your first MCP server, then a client, then a client backed by an LLM.
- Switch from stdio to streamable HTTP, add authentication, and deploy a server.
- Debug a live server with the MCP Inspector and connect popular host clients to it.
- Work through the security module before you expose anything to the internet.
On LearnOS you study it with notes, progress tracking and Q&A that cites the lesson section it drew from.