sourcebolt

sourcebolt

Let your AI assistant use your Google NotebookLM notebooks: one small, fast program, nothing else to install.

Install in 2 minutes

You want to...	Go here
Run one command that installs and connects everything	The two commands below
Follow pictures for your AI app (Claude, Cursor, VS Code, Windsurf, Gemini CLI, ChatGPT)	Step-by-step guides
Let your AI app install it for you	Let your AI install it
Use npm, Docker or a manual setup (developers)	Install and Connect your AI app

Windows (open PowerShell, paste, press Enter):

irm https://raw.githubusercontent.com/candidloom-hub/sourcebolt/main/install.ps1 | iex

macOS and Linux (open Terminal, paste, press Enter):

curl -fsSL https://raw.githubusercontent.com/candidloom-hub/sourcebolt/main/install.sh | sh

The installer checks the download, installs sourcebolt for your user only, and then offers sourcebolt setup. Setup finds your AI apps, connects the ones you pick (keeping a backup of their settings), and opens a browser window so you can sign in to Google. Your password goes only to Google.

Quick start · Install · Connect your app · Tools · FAQ · Docs

Independent project. sourcebolt is not made by, endorsed by or sponsored by Google. "NotebookLM" is a trademark of Google LLC. NotebookLM has no public API for personal accounts, so sourcebolt uses the same internal web calls the NotebookLM website makes. Read Limits and risk before you rely on it.

What it does

sourcebolt is an MCP server. MCP (Model Context Protocol) is the standard way AI apps such as Claude, Cursor, VS Code and Codex call outside tools. Once sourcebolt is connected, you can ask your assistant things like:

• "List my NotebookLM notebooks."
• "Add these three articles to my Thesis notebook."
• "Ask the Thesis notebook what the sources say about sample size, with citations."
• "Research recent work on battery recycling and import the best sources."

The assistant calls sourcebolt, sourcebolt talks to NotebookLM with your own Google session, and the answer comes back with citations from your sources.

Key numbers

Measured on the GitHub Actions runner (2 CPUs, AMD EPYC 7763), v0.2.0, medians. Method and raw data: bench/RESULTS.md and docs/performance.md.

What	Result
Start-up time (start to ready)	8.2 ms
Program size	about 10 MB, one file
Memory while idle	16.2 MB
sourcebolt's own work per call, 100 notebooks, cached / not cached	0.40 ms / 0.76 ms
Chat: first progress after Google's first chunk	+0.3 ms
A real notebook_list call to Google	848 ms median (Google's time dominates; one laptop, 2026-10-14)

Quick start

1. Install. Download the file for your system from the latest release and put sourcebolt on your PATH. Step-by-step commands are in Install.

2. Sign in. Start Brave or Chrome with a debugging port, sign in to NotebookLM in that window, then run:

   sourcebolt login --attach http://127.0.0.1:9222
   

   You should see Saved profile "default" for notebook.google.com (N notebooks visible). The browser commands for each system are in docs/auth.md. From the release after v0.2.0, sourcebolt setup does steps 2 to 4 for you (docs/getting-started.md).

3. Check it.

   sourcebolt doctor
   

   Every line should start with ok.

4. Connect your AI app. For Claude Code:

   claude mcp add sourcebolt -- sourcebolt serve
   

   After v0.2.0 you can instead install the Claude Code plugin, which adds the server and the sourcebolt skill (the program must still be on your PATH):

   /plugin install sourcebolt --marketplace candidloom-hub/sourcebolt
   

   Other apps: Connect your AI app. One-click install links: docs/install-links.md.

5. Ask a question. In your app, type: "Use sourcebolt to list my NotebookLM notebooks, then ask the first one for a one-paragraph summary with citations."

The full walkthrough with expected output is in docs/getting-started.md.

Install

The current release is v0.3.3. Not sure which way to pick? See docs/install.md, or use the one-line installer (it verifies the download's checksum, needs no admin rights, and -Uninstall / --uninstall undoes it):

irm https://raw.githubusercontent.com/candidloom-hub/sourcebolt/main/install.ps1 | iex   # Windows PowerShell

curl -fsSL https://raw.githubusercontent.com/candidloom-hub/sourcebolt/main/install.sh | sh   # macOS, Linux

Installing through an AI agent such as Cline? Point it at llms-install.md: agent-ready steps, where you still sign in to Google yourself.

Channel	Status	How
GitHub release binary	Available	Windows, macOS and Linux, on x64 and ARM64. See below.
go install	Available	Needs Go 1.27 or newer.
Docker (GHCR)	Available	ghcr.io/candidloom-hub/sourcebolt:0.3.3. Images are tagged by version only; there is no latest tag.
npm (npx)	Available	npx -y sourcebolt serve (needs Node.js 18+). In an app's MCP config: command npx, args ["-y", "sourcebolt", "serve"].
PyPI (pipx, uvx)	Not published yet

Release binary

Windows (PowerShell). This puts sourcebolt.exe in a folder in your user profile and adds that folder to your PATH. For an ARM64 PC, replace amd64 with arm64.

$dir = "$env:LOCALAPPDATA\Programs\sourcebolt"
New-Item -ItemType Directory -Force $dir | Out-Null
Invoke-WebRequest https://github.com/candidloom-hub/sourcebolt/releases/download/v0.3.3/sourcebolt_0.3.3_windows_amd64.zip -OutFile "$env:TEMP\sourcebolt.zip"
Expand-Archive "$env:TEMP\sourcebolt.zip" -DestinationPath $dir -Force
[Environment]::SetEnvironmentVariable("Path", [Environment]::GetEnvironmentVariable("Path", "User") + ";$dir", "User")

Open a new terminal, then run sourcebolt --version. It prints sourcebolt 0.3.3.

macOS and Linux (bash). Set OS to darwin or linux and ARCH to amd64 or arm64.

OS=linux ARCH=amd64
curl -fsSLO "https://github.com/candidloom-hub/sourcebolt/releases/download/v0.3.3/sourcebolt_0.3.3_${OS}_${ARCH}.tar.gz"
tar -xzf "sourcebolt_0.3.3_${OS}_${ARCH}.tar.gz" sourcebolt
mkdir -p ~/.local/bin && mv sourcebolt ~/.local/bin/
sourcebolt --version

If sourcebolt is not found, add ~/.local/bin to your PATH. Each release has a checksums.txt with a keyless cosign signature and a CycloneDX SBOM (software bill of materials) per archive; docs/security.md shows how to check them.

Go

go install github.com/candidloom-hub/sourcebolt/cmd/sourcebolt@latest

Docker

docker run -i --rm -v sourcebolt-config:/config ghcr.io/candidloom-hub/sourcebolt:0.3.3

The image has no OS keyring, so the profile key is kept in a file inside the /config volume. Protect that volume like a password. The 0.2.0-browser tag adds a browser for the fallback path. Signing in inside a container: docs/auth.md.

Connect your AI app

Sign in first. All apps start the same command: sourcebolt serve. If sourcebolt is not on your PATH, use the full path to the program instead (on Windows, for example C:\\Users\\you\\AppData\\Local\\Programs\\sourcebolt\\sourcebolt.exe inside JSON).

App	Works locally?	Setup
Claude Code	Yes	claude mcp add sourcebolt -- sourcebolt serve, or after v0.2.0 the plugin: /plugin install sourcebolt --marketplace candidloom-hub/sourcebolt
Claude Desktop	Yes	Add to claude_desktop_config.json (below)
Cursor	Yes	Add to ~/.cursor/mcp.json (below)
VS Code (Copilot agent mode)	Yes	code --add-mcp '{"name":"sourcebolt","type":"stdio","command":"sourcebolt","args":["serve"]}'
Windsurf	Yes	Add to its mcp_config.json (below)
Codex CLI and IDE	Yes	codex mcp add sourcebolt -- sourcebolt serve
Gemini CLI	Yes	gemini mcp add -s user sourcebolt sourcebolt serve
ChatGPT desktop app	Yes, as a local (STDIO) server	Settings, MCP servers; it shares the Codex configuration
ChatGPT on the web	No	Needs a public HTTPS server. sourcebolt runs only on your computer, by design.

The JSON format used by Claude Desktop, Cursor and Windsurf:

{
  "mcpServers": {
    "sourcebolt": { "command": "sourcebolt", "args": ["serve"] }
  }
}

File locations, VS Code JSON, Codex TOML, the HTTP mode and which snippets were tested: docs/clients.md.

Small tool budget? Add "--tools", "minimal" to args to expose only the 8 core tools.

What your assistant can do

Grouped by task. Every tool, option and state is in the generated reference, docs/tools.md.

Task	Tools
Find and open notebooks	notebook_list, notebook_open, notebook_create, notebook_update
Add and read sources (web pages, YouTube, pasted text)	source_add, source_list, source_read, source_update
Ask questions with citations	chat_ask, chat_status, chat_history, chat_configure, suggest
Find new sources on the web	research_start, research_status, research_import
Keep notes	note_list, note_save
Studio (audio, reports, flashcards, ...)	studio_list, studio_generate
Sharing, labels, account	share_status, share_update, label_list, label_update, account_settings
Clean up	delete (always previews first), job_cancel

Some options are not available yet, for example Drive and file uploads and the Studio option values no live capture has confirmed. They return a clear UNSUPPORTED_BY_TRANSPORT error instead of failing silently. SKILL.md tells an agent which tool to use when.

How it works

1. Direct calls first. sourcebolt sends NotebookLM's own internal web requests (called RPC, remote procedure calls) over HTTP/2 with your session cookies. This is the fast path.
2. Browser fallback. If Google rejects the direct request, sourcebolt can repeat it inside a headless Chrome-family browser on a separate profile. It needs Brave, Chrome or Edge installed (or the -browser Docker image).
3. Page fallback. If Google changes the format of the chat or add-source request, sourcebolt can type into the NotebookLM page itself and read the answer (the "DOM" fallback, named after the page structure it drives).

Turn the browser paths off with --no-browser-fallback. server_info shows which path served the last call. Details: docs/auth.md.

FAQ

Is this an official Google product? No. It is an independent open-source project.

Does sourcebolt see my Google password? No. You sign in yourself, in a normal browser window. sourcebolt keeps only the session cookies, encrypted on your disk.

Can the AI read my cookies? No tool accepts or returns cookies or passwords. Sign-in works only from the command line.

Does it send my data anywhere else? No. It talks only to NotebookLM. URLs you add as sources are fetched by Google, not by your computer.

Is it free? Yes. Apache-2.0, no paid tier.

More answers: docs/faq.md. Something broken? docs/troubleshooting.md.

Limits and risk

• Unofficial. Google can change NotebookLM's internal calls at any time. Tools may break until sourcebolt is updated.
• Terms of Service. Automating a Google web app with your session may break Google's terms and could lead to rate limits or account action. You use sourcebolt at your own risk; there is no warranty (see the license).
• Google's quotas apply. Limits on notebooks, sources, chats and Studio are set by Google per account. sourcebolt reports them as errors and does not bypass them.
• Try it safely first. Use a test notebook with public content while you evaluate it.

Documentation

Start at docs/index.md. Highlights:

• Step-by-step guides per app · Let your AI install it
• Getting started · Connect your AI app · Troubleshooting · FAQ
• Tool reference · Sign-in and storage · Security · Privacy · Performance
• One-click install links · Registries and directories
• Why sourcebolt · Comparison with other NotebookLM servers · Changelog

Contributing

Bug reports and pull requests are welcome on GitHub Issues. Before you open one:

• Run sourcebolt --version and sourcebolt doctor and include the output. Remove any cookies or tokens first.
• For code changes, run go vet ./... and go test ./....
• docs/tools.md is generated. Change the tool descriptions in code, then run go run ./tools/gendocs > docs/tools.md.
• Report security problems privately, as described in docs/security.md.

License

Apache-2.0. See LICENSE and NOTICE.

Credits

sourcebolt is written from scratch. No source code from other projects was copied. Designs and knowledge of the NotebookLM wire format were learned, with thanks, from these MIT- or Apache-2.0-licensed projects (see NOTICE and research/notes/). Projects with no license or a copyleft license were read for behavior only, and nothing from them was used.

Project	License	What we learned from it
teng-lin/notebooklm-py	MIT	RPC ids, payload layouts, source and chat flows
tmc/nlm	MIT	Go wire-protocol modeling and lossless round-trip testing
jacob-bd/gemini-notebook-mcp-cli	MIT	MCP tool surface, internal API reference
LocalKinAI/notebooklm-go	Apache-2.0	Go client and auth flow ideas