Ever since I started highlighting text on my Kindle, then my PocketBook, and also the old-fashioned way with pen and paper, I have been trying to store those highlights in my digital garden — which serves as a second brain for recalling insights from books I have read. I built brainfck.org to keep my collection of notes available online, not only for myself but also for others. I even added a search feature (first introduced in ce36502 and later improved in 9a23b40) that lets you search across the entire collection. Whenever I want to look something up, I go to brainfck.org, type in a keyword, and it shows me the books and journal entries where that word appears.

Recently I built a workflow that allows to expand all this. I now have all my e-books loaded into NotebookLM notebooks, organised year by year, so I can ask questions across everything I have read in a given period. The process is to convert EPUBs and PDFs to Markdown, upload them into the relevant notebook, and start asking questions. A particularly useful feature is that you can select a single source and have a focused conversation about just that one book.

This post describes how I built it using Claude Code with the NotebookLM MCP server and the Dropbox (where I usually store my EPUB files).

The goal #

The end state is simple: one NotebookLM notebook per year (Books 2020, Books 2021, …, Books 2026), each containing the full text of every book I read that year. With that in place I can do things like:

“What did these books say about motivation and focus?”

“Summarise each book in three bullet points”

“Which books overlap on the topic of climate change?”

NotebookLM notebooks overview

The tools #

I used three MCP servers inside Claude Code:

• notebooklm-mcp — create notebooks, add sources, query them
• DropboxMCP — search and download EPUBs/PDFs from my PocketBook library synced to Dropbox
• WebFetch — scrape brainfck.org/books to get the canonical list of books per year

To keep MCP servers scoped per repository, add them to your project’s .claude.json under mcpServers:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

{
  "mcpServers": {
    "notebooklm-mcp": {
      "type": "stdio",
      "command": "notebooklm-mcp"
    },
    "DropboxMCP": {
      "type": "http",
      "url": "https://mcp.dropbox.com/mcp"
    }
  }
}

Install the NotebookLM CLI once globally with:

1

uv tool install notebooklm-mcp-cli

The Dropbox MCP is a hosted HTTP server — no local install needed, just authenticate once via claude mcp auth DropboxMCP in the project directory.

The workflow #

Step 1: Get the book list for a year #

The brainfck books page lists the books I have read, sorted by year. Starting with only two notebooks (“Books 2026” and “Books 2025”), I wanted to create notebooks for earlier years but I did not know which books I actually had as EPUBs in Dropbox.

brainfck.org/books

1
2

WebFetch https://brainfck.org/books/
→ returns books grouped by year with their read dates

With the book list in hand, I also asked Claude to list the existing NotebookLM notebooks so I could see which years were already covered and how many sources each had:

Claude Code listing NotebookLM notebooks

Step 2: Search Dropbox for the files #

For each book I ran a parallel Dropbox search:

I asked Claude to search for each book title using the DropboxMCP MCP tool. Under the hood, Claude Code calls the tool directly:

1
2
3
4
5

# Claude Code MCP tool call (one per book, all run in parallel)
mcp__DropboxMCP__search
  query: "Book Title"
  file_extensions: ["epub", "pdf"]
  filename_only: true

Claude ran all searches in parallel, then cross-referenced the results against the book list. Many books lived in year-based folders (E-Books/2024/, E-Books/2025/, etc.), while older ones were in the root E-Books/ folder or the Golang/ subfolder.

Step 3: Extract text content #

For files under 5MB, the Dropbox MCP can extract text directly:

1
2

mcp__DropboxMCP__get_file_content
  path_or_file_id: "id:aTF8BX-..."

For larger files (5–15MB EPUBs, oversized PDFs), I had to generate shared download link:

1
2
3

mcp__DropboxMCP__create_shared_link
  path_or_file_id: "id:aTF8BX-..."
  allow_download: true

Then downloaded locally and converted with pandoc:

1

pandoc -t markdown book.epub -o book_full.md

For PDFs, pdftotext worked better:

1

pdftotext book.pdf book_full.md

Step 4: Create the notebook and upload #

With the converted Markdown files in hand, I first created a new notebook for the year, then added each book as a source. The notebook_create call returns a notebook ID which is then passed to every subsequent source_add call:

1
2
3
4
5
6
7
8
9

# 1. Create the notebook for the year (returns a notebook_id)
mcp__notebooklm-mcp__notebook_create title: "Books 2024"

# 2. For each book: upload the converted Markdown file as a source
mcp__notebooklm-mcp__source_add
  notebook_id: "..."          # ID returned from step 1
  source_type: file
  file_path: "/path/to/book_full.md"
  title: "The Effective Executive"

All source_add calls for a given year ran in parallel — typically 6–8 at once — so populating a full year’s notebook took only a minute or two.

Querying across years #

With all notebooks populated, I can now query a year’s reading. But even simpler — I can just ask Claude Code in plain English:

Asking Claude Code about Books 2026

Or go deeper with a structured query:

1
2
3

mcp__notebooklm-mcp__notebook_query
  notebook_id: "Books 2026"
  query: "Give me a short summary of each book — main argument and key takeaways"

The answers are grounded in the source material and surprisingly detailed. Here’s an example querying the main concepts from Earth for All:

Querying Earth for All via Claude Code

This is particularly useful for rediscovering insights from books I read years ago and for finding thematic threads across multiple books in the same period.

Inside the Books 2025 notebook

The whole setup took a few hours across multiple sessions — most of that was handling edge cases (DRM, oversized files, misplacements). The actual notebook creation and uploading per year takes about 5 minutes once the files are in order.