Open Notebook describes itself as “An open source, privacy-focused alternative to Google’s Notebook LM!” — you bring your own documents and your own model, and it runs on your hardware. It is MIT licensed, at version 1.15.0, and actively developed.
The word doing the most work in that description is “privacy-focused,” and it is conditional in a way worth being precise about.
Privacy is a configuration, not a property
Self-hosting Open Notebook means your documents sit on your disk and your notebook is not someone else’s product. That is real and it is the main reason to use it.
It does not, by itself, mean your documents stay on your machine. Open Notebook is a front end for language models, and the README references OpenAI, Anthropic, Groq, Mistral, DeepSeek, Vertex, OpenRouter and ElevenLabs alongside Ollama. If you configure it with an OpenAI key, your source material is sent to OpenAI when you ask a question about it. The application is private; the inference is wherever you pointed it.
So the honest version of the pitch is: Open Notebook gives you the choice that Google’s product does not. Taking the private option means Ollama, running models on your own hardware, and that is the configuration to aim for if privacy is why you came.
Be realistic about what that costs. Local models need RAM and ideally a GPU, and a small local model summarising a dense PDF is noticeably worse than a frontier model doing the same job. That trade is the actual decision, and it is yours rather than the software’s. A reasonable middle path many people land on: Ollama for anything confidential, a hosted provider for public papers and general research, configured side by side.
What the product genuinely guarantees regardless of provider is that nobody is training on your notebook as a condition of using it, and that you can take your documents and leave.
It brings an unusual database
The deployment is two containers, and the second one is the part to notice:
surrealdb:
image: surrealdb/surrealdb:v2
open_notebook:
image: lfnovo/open_notebook:v1-latestSurrealDB, not Postgres or SQLite. It is a capable multi-model database and it is also a young and comparatively uncommon one, which has practical consequences you should price in before you put a year of research notes behind it:
- Backup and restore tooling is less familiar. You will not be reaching for
pg_dumpand the thirty years of operational knowledge behind it. Work out your backup procedure and — more importantly — test a restore before you rely on it. - Fewer people can help you. When something goes wrong at the database layer, the pool of answers is smaller.
- Pin the tag. The compose file names
surrealdb:v2, which is a major-version track rather than an exact version. For a datastore holding your notes, pin the specific version and upgrade deliberately.
The application image is v1-latest, which has the same consideration: convenient for evaluation, and a moving target in production.
None of this is a reason to avoid Open Notebook. It is a reason to treat the data volume as the asset and to prove you can get the data back out. Python 3.11–3.12 is the stated requirement if you run it outside Docker.
What the repository tells you about its seriousness
The root of the project is unusually well furnished for a tool of this size: VISION.md, CONFIGURATION.md, SECURITY.md, MAINTAINER_GUIDE.md, CONTRIBUTING.md, a CODE_OF_CONDUCT.md, a separate README.dev.md, a CHANGELOG.md and a structured docs/ tree with numbered sections from “START HERE” through installation, core concepts and a user guide.
That is a project that expects other people to contribute and to operate it, rather than one person’s working directory published. At 647 tracked files and version 1.15.0, it is also past the stage where a README is the documentation.
A VISION.md is the one worth reading before you adopt it. For a tool in a category defined by a Google product, where the project intends to go matters more than its current feature list.
What it is for
The Notebook LM shape is specific and genuinely useful: you load a set of sources — papers, reports, transcripts, notes — and then interrogate that set rather than the open web. The answers are grounded in your documents, and the value is that it will not confidently cite something you never gave it.
Open Notebook fits the person doing that with material they would rather not upload: client documents, unpublished research, internal strategy, medical or legal papers, a thesis bibliography. If the sources are public papers you do not mind sharing, Google’s product is less work.
For a durable store of your own thinking rather than a question-answering layer over documents, the comparison is Obsidian — plain Markdown files on disk, with no database to operate. Many people end up using both, and getting documents into a form either tool can read is where a converter like MarkItDown earns its place.
How to start
Run it with Ollama first, even if you intend to use a hosted model later. It takes the same effort and it shows you the fully private configuration working, so you know the fallback exists. Load five documents rather than five hundred. Then back up the SurrealDB volume and restore it into a fresh instance — do that once, early, while losing the data would not matter.
The finding: Open Notebook’s privacy claim is a choice it offers rather than a guarantee it makes, and realising it means Ollama and accepting weaker local models for sensitive material. The operational detail nobody mentions is SurrealDB — a young database holding your research notes, pinned to a major-version tag in the sample compose file. Pin it properly and test a restore before you trust it.
Primary link
Learn more at: https://github.com/lfnovo/open-notebook
