NotebookLM, but the data and the models are yours
Open Notebook rebuilds the idea behind Google’s NotebookLM as software you host yourself: pull in PDFs, videos, audio, and web pages, search across them, chat with that corpus as context, and generate a podcast-style audio summary. The pitch is not “we cloned the UI.” It is that two things Google keeps closed, your data and your choice of model, are open here.
That framing is the whole reason to run it. If your research is sensitive, or you want to point it at a cheaper or local model instead of a single vendor’s, NotebookLM cannot do that and Open Notebook is built for exactly it. Under the hood it pairs a Python backend with a Next.js and React frontend, stores everything in SurrealDB, and orchestrates models through LangChain.
What you get over the hosted original
- Model choice across 18+ providers, including OpenAI, Anthropic, Ollama, and LM Studio, so you can run fully local or mix vendors by cost.
- Multi-modal sources: PDFs, video, audio, and web pages organized into notebooks.
- Podcast generation with 1 to 4 speakers and custom speaker profiles, against NotebookLM’s fixed two-host format.
- Full-text and vector search across everything you have loaded.
- A full REST API, which the hosted product simply does not expose, so you can automate ingestion and generation.
- Deploy anywhere: Docker, cloud, or a local machine, with a multi-language UI.
The workflow it is built around
The pieces fit into a research loop rather than a pile of features. You load sources into a notebook, optionally run content transformations over them, custom or built-in, to reshape raw material into summaries or structured notes, then search across everything with combined full-text and vector retrieval. From there you chat with that corpus as grounded context, and when you want a passive review you generate a multi-speaker podcast from it. The transformations step is the part most NotebookLM clones skip, and it is where the open model earns its flexibility: you decide how content is processed rather than accepting a fixed pipeline.
Install
The documented path is Docker Compose, and it expects Docker Desktop and nothing else up front; API keys are entered in the UI later. Grab the compose file and bring it up:
curl -o docker-compose.yml https://raw.githubusercontent.com/lfnovo/open-notebook/main/docker-compose.yml
docker compose upThe compose file runs SurrealDB alongside the app, so the database is provisioned for you rather than being a separate install. From there you configure providers in the interface.
Where it honestly trails NotebookLM
This is the part worth dwelling on, because the project is candid about it. Its own comparison table marks citations as “basic references (will improve),” while listing NotebookLM’s as “comprehensive with sources.” For a research tool, source-grounded citations are not a side feature, they are the trust layer. If your work depends on tracing every claim back to the exact passage, the hosted original is currently stronger on that axis, and you should weigh that against the privacy and flexibility you gain.
The tracker reflects an actively evolving project rather than a settled one: 161 open issues as of 2026-06, with frequent releases (v1.9.0 in June 2026). Treat citation fidelity and edge-case stability as moving targets you should test against your own corpus before committing a workflow.
Open Notebook versus NotebookLM
| Open Notebook | Google NotebookLM | |
|---|---|---|
| Hosting | self-hosted, your data | Google cloud only |
| Models | 18+ providers, local or hosted | Google models only |
| Podcast voices | 1 to 4, custom profiles | 2 fixed hosts |
| API | full REST API | none |
| Citations | basic, improving | comprehensive with sources |
| License | MIT | proprietary |
This comparison follows the project’s own documented feature matrix. NotebookLM is a closed hosted product with no public repository to measure by stars, so the contrast is on capability, not popularity. The trade is consistent: Open Notebook wins on control, flexibility, and automation; NotebookLM currently wins on citation depth and zero-setup polish.
Related
Feed it cleaner inputs by converting documents first with MarkItDown, and run it against a local model via Ollama to keep everything on your own machine. For what else is trending, see the daily digest and the weekly report.
FAQ
Is Open Notebook free? The software is MIT-licensed and self-hosted. Your only cost is whatever AI provider you connect, and you can run locally for none.
Can I run it fully offline? You can point it at local models through Ollama or LM Studio, keeping inference on your machine. The app and database run in Docker locally.
How is it different from NotebookLM? Self-hosting, 18+ model providers, 1 to 4 podcast speakers, and a REST API. The main trade-off is weaker citations today.
What database does it use? SurrealDB, provisioned automatically by the provided Docker Compose file.