A free, open-source, offline alternative to Google NotebookLM for macOS and Windows. Chat with your own PDFs, Office documents, and web pages and get answers with inline citations — privacy-first and local-first, with everything running on your machine. Powered by Ollama; inspired by open-notebook.
⬇️ Download the latest release — macOS .dmg · Windows .exe
If AI Notebook is useful to you, please ⭐ star the repo — it helps others find it.
AI Notebook is a native desktop app for grounded research over your own
sources. Drop in PDFs, Office documents, web pages, and plain text; the app
chunks and embeds everything locally, then lets you ask questions and get
answers with inline [N] citations that link straight back to the exact
snippet they came from.
The core idea: a NotebookLM-style workflow where your documents never leave your machine unless you explicitly choose a cloud model. Embeddings, retrieval, notes, and the database are all local. On macOS the AI is local-only (Ollama). On Windows you can stay fully local with Ollama, or opt into a cloud provider (Anthropic, OpenAI) when you want a stronger model.
It ships as two native codebases sharing one design and data model:
| Platform | Stack | AI providers |
|---|---|---|
| macOS 14+ | Swift 6 · SwiftUI | Ollama (local only) |
| Windows 10/11 | .NET 10 · WinUI 3 | Ollama (local) · Anthropic · OpenAI · any OpenAI-compatible endpoint |
| AI Notebook | Google NotebookLM | |
|---|---|---|
| Where it runs | Your machine (native desktop app) | Cloud / browser |
| Your documents | Never leave your device (with Ollama) | Uploaded to Google |
| Works offline | ✅ Yes | ❌ No |
| Cost | Free & open source (MIT) | Free tier, Google account required |
| AI model | Local (Ollama) or bring your own key | Google models only |
| Platforms | Native macOS & Windows | Web only |
If you want NotebookLM-style "chat with your sources, get citations" but without sending your files to the cloud, AI Notebook is built for that.
macOS app, showing the "Renewable Energy" demo notebook.
| Sources — local ingestion & indexing | Chat with [N] citations |
|---|---|
 |
 |
| Notes — WYSIWYG editor + note-scoped chat | AI tools — transformation templates |
 |
 |
- Notebooks — organise research into separate projects.
- Sources — ingest PDF, TXT, Markdown, web URLs, and Word / PowerPoint / Excel. Each source is chunked and embedded locally.
- Chat with citations — ask questions across your sources; answers stream
in with clickable
[N]chips that reveal the cited snippet. Scope a chat to a specific subset of sources when you want a narrow answer. - Hybrid retrieval — semantic search (cosine similarity over local embeddings) fused with keyword search (SQLite FTS5 / BM25) via Reciprocal Rank Fusion, so both meaning and exact terms are matched.
- Notes — a TipTap WYSIWYG editor with autosave, version history, and file attachments. Write by hand or save any AI answer as a note. Notes are auto-indexed and searchable alongside sources.
- Transformations ("AI tools") — run a prompt template over a source and store the result as a note. Built-in templates (Summary, Key points, Entities) plus your own custom prompts, with batch runs and history.
- Follow-up suggestions & per-source summaries — the chat surfaces suggested follow-up questions and one-line summaries for each source.
- Multi-provider AI (Windows) — route chat and embeddings to Ollama, Anthropic, OpenAI, or any OpenAI-compatible server. Keys are stored in the Windows Credential Manager, never in the database.
- Optional web search (Windows) — opt-in web results injected as user-message context (kept out of the system prompt to limit prompt injection).
- Export / import / backup (Windows) — export a single note to Markdown or a whole notebook to a ZIP archive; back up and restore the full database.
- Global search, tags & source sets (Windows) — search across the whole notebook, tag sources, and save reusable source sets for scoped chat.
- Bilingual UI — English and Czech, auto-detected from the system locale and switchable in Settings.
Platform parity: the macOS app focuses on a tight local-only Ollama workflow. The Windows app additionally ships multi-provider AI, web search, export/backup, tags, source sets, and global search.
Everything runs on your machine by default:
- Source text, chunks, embeddings, notes, and chat history live in a single local SQLite database.
- API keys (Windows cloud providers) are stored in the OS credential vault, not in the database or in plain text.
- The only outbound network calls are ones you initiate: fetching a web URL you add as a source, an optional update check, optional web search, and — on Windows — requests to a cloud AI provider you explicitly configure.
- With Ollama as the provider, the app makes no AI calls off your device.
Download the latest AINotebook-vX.Y.Z-macos.dmg from
Releases, open it, drag
AI Notebook to Applications, and launch. The first run walks you through
Ollama detection and model download.
Requirements
- macOS 14 (Sonoma) or later
- Ollama (the app guides you through setup on first launch)
- ~5 GB free disk for the default models (
llama3.2:3b+nomic-embed-text)
Download the latest AINotebook-vX.Y.Z-windows-setup.exe from
Releases and run the
installer. The build is self-contained; the WebView2 runtime is installed
automatically if missing.
Requirements
- Windows 10 (1809+) or Windows 11
- For local AI: Ollama. For cloud AI: an API key for your chosen provider (Anthropic / OpenAI / OpenAI-compatible).
The release binaries are not signed with a paid Apple/Microsoft developer certificate, so the OS shows a warning the first time you open the app. This is expected for an open-source build — you can verify the source here and build it yourself. To run the downloaded release:
macOS — Gatekeeper blocks it with "AI Notebook can't be opened because Apple cannot check it for malicious software" (or "unidentified developer").
- In Finder, right-click (or Control-click) AI Notebook in Applications → Open, then confirm Open in the dialog. macOS remembers the choice for next time.
- If there's no Open button, go to System Settings → Privacy & Security, scroll to the message about AI Notebook, and click Open Anyway.
- Stubborn quarantine flag? Clear it in Terminal:
xattr -dr com.apple.quarantine "/Applications/AI Notebook.app"
Windows — SmartScreen shows "Windows protected your PC" with "Unknown publisher".
- Click More info.
- Click Run anyway to launch the installer.
- If a User Account Control prompt names an unknown publisher, confirm Yes to continue.
The releases aren't signed with a paid certificate, but every release artifact carries a GitHub build-provenance attestation (signed via Sigstore, recorded in a public transparency log). It cryptographically proves the file was built by this repo's release workflow and not tampered with. With the GitHub CLI:
gh attestation verify AINotebook-v0.8.1-macos.dmg --repo lukoplt/AI-notebook
# or, on Windows:
gh attestation verify AINotebook-v0.8.1-windows-setup.exe --repo lukoplt/AI-notebookA successful check confirms the binary's origin. (This does not remove the OS first-launch warning above — only a paid Apple/Microsoft certificate does that.)
git clone https://github.com/lukoplt/AI-notebook
cd AI-notebook
swift run AINotebookAppRequires Xcode 16+ (Swift 6).
git clone https://github.com/lukoplt/AI-notebook
cd AI-notebook
dotnet restore windows/AINotebook.sln --locked-mode
dotnet build windows/AINotebook.sln -c ReleaseRequires the .NET 10 SDK and the Windows App SDK / WinUI 3 workload.
NuGet dependencies are pinned via packages.lock.json — restore in locked
mode so a drifting lockfile fails the build. See
docs/windows-build.md for packaging the installer.
A shared design and data model implemented natively per platform.
macOS (Sources/)
AINotebookCore— Swift Package library: SQLite storage (GRDB), Ollama client, ingestion, embedder, hybrid retriever, chat engine, transformations.AINotebookApp— SwiftUI executable.- Database:
~/Library/Application Support/AINotebook/db.sqlite.
Windows (windows/)
AINotebook.Core— .NET class library: extractors, ingestion, RAG pipeline, Dapper/SQLite storage, Ollama client, and the multi-provider router (Anthropic, OpenAI, OpenAI-compatible) with a Credential-Manager-backed key store.AINotebook.App— WinUI 3 app (WebView2-hosted TipTap editor).
Shared model — 11 tables with a versioned migration history, hybrid retrieval (cosine + FTS5 BM25 → Reciprocal Rank Fusion), and a localization layer (EN/CZ, ~155 keys) common to both platforms.
The full design spec lives in
docs/superpowers/specs/; the development roadmap is
in docs/roadmap.md.
Sources/ macOS — Swift / SwiftUI app + AINotebookCore library
Tests/ macOS — Swift unit tests
windows/ Windows — .NET 10 / WinUI 3 solution
src/AINotebook.Core shared-design core (storage, RAG, providers)
src/AINotebook.App WinUI 3 application
tests/ .NET unit tests
installer/ Inno Setup script + redistributables
docs/ design spec, roadmap, build notes
.github/workflows/ CI + release pipelines (macOS & Windows)
Issues and pull requests are welcome. Both platforms are covered by CI
(core-ci, windows-ci) on every push — please make sure tests pass and, on
Windows, that the NuGet lockfile stays in sync (--locked-mode).
Inspired by open-notebook (MIT). Built with GRDB.swift, SwiftSoup, ZIPFoundation (macOS) and AngleSharp, Dapper, PdfPig, WebView2, Windows App SDK (Windows). Local AI powered by Ollama.
Made with ❤️ by Lukáš Oplt.