Cortex

A local-first, open-source NotebookLM alternative — a desktop study OS for serious learners.

Ingest everything a course throws at you — slides, PDFs, docs, lecture recordings, web pages, YouTube — into a clean Subjects → Topics → Sources hierarchy, then turn it into exam-ready study material: cheatsheets, flashcards with spaced repetition, quizzes, two-host audio overviews, infographics, mind maps, and a citation-grounded chat.

Local-first and bring-your-own-keys: your sources, embeddings, and generated study material live in a local database on your machine. The only data that leaves it is the content you send to the AI provider you configure (cheatsheets, chat, quizzes, etc. are generated by that provider) and, optionally, your own SearXNG.

Why Cortex? · Preview · Features · Download · Build from source · Configuration · How it works · Roadmap · Contributing

🌐 Website · 📖 Docs · 🔒 Privacy Policy · 💬 Support

---

Preview

A study-analytics dashboard that actually earns the name — accuracy, focus-by-hour, consistency streaks, and per-topic mastery.

Auto-generated cheatsheets — exam-focused, structured study notes synthesised per topic, with tables, callouts, formulas and charts.	Citation-grounded chat — ask questions across a subject and get answers backed by your own sources.
Spaced-repetition flashcards — FSRS scheduling with Again / Hard / Good / Easy and visible intervals.	Generated study material — flashcards, quizzes, audio overviews, infographics and mind maps from your sources.
Calendar & assignment tracking — deadlines, exams and a kanban planner, colour-coded by subject.	Everything organised — PDFs, slides, docs, web pages and YouTube ingested into Subjects → Topics → Sources.

Record lectures with live transcription — straight into the right subject.

🎨 Every screenshot above is a different built-in theme. Cortex ships a range of themes (Everforest, Gruvbox, Nord, Dracula, Rosé Pine, Kanagawa, Solarized…) and can even follow your Omarchy desktop theme automatically.

---

Why Cortex?

NotebookLM is great, but it's a web product: your sources live on someone else's servers, you can't pick your model, and it's organized around loose "notebooks." Cortex is the opposite:

	NotebookLM	Cortex
Data location	Cloud	100% local SQLite on your machine
AI model	Fixed	Bring your own — Gemini, OpenRouter, OpenAI, Claude, or local Ollama
Structure	Flat notebooks	Subjects → Topics → Sources
Study materials	Audio + notes	Cheatsheets, FSRS flashcards, quizzes, timed graded exams, audio, infographics, mind maps, slides
Citations & deadlines	—	Built-in reference manager (APA/MLA) + exam/assignment tracking
Export	—	Anki .apkg, PDF, portable SQLite, encrypted backups
Cost	Subscription	Free & open source (you pay only your own API usage)

Features

Ingest anything

• Sources: PDF, PPTX, DOCX, plain text/Markdown, web pages, YouTube, audio recordings, and images via vision-model OCR — including photographed handwritten notes.
• Office docs are rendered to PDF for faithful slide previews; PDFs use real text extraction.
• Each source is parsed → chunked → embedded → stored with live progress, then becomes searchable and citable.

Generate study material

• Cheatsheets — exhaustive, exam-focused, completeness-checked synthesis with callouts, tables, and bar charts. Optional web-sourced diagrams per section, with versioned drafts and an approve-diff review flow.
• Flashcards with FSRS spaced repetition (the modern memory-model scheduler; Again/Hard/Good/Easy, "Study due · N").
• Exam mode — timed mock papers (MCQ + written) generated from your sources; MCQs grade locally, written answers are graded by your model with a verify-before-score rubric, per-question feedback, weak-topic callouts, and a one-click remark that re-grades with the identical rubric.
• Quizzes — multiple-choice with explanations.
• Audio overviews — two-host, podcast-style spoken walkthroughs, rendered to real audio via cloud TTS (on-device speech synthesis as fallback).
• Infographics — detailed HTML posters with a timeline of events, key stats, and a takeaway.
• Mind maps — hierarchical concept maps.
• Slideshows — presentation outlines.
• Every generator accepts an optional custom prompt (NotebookLM-style) to steer focus and tone.

Chat that cites its sources

• Ask questions scoped to a subject, topic, a tag, or specific sources; answers cite inline as ⟦source · location⟧.
• Web mode 🌐 pulls in live web results and, for visual questions, fetches images and diagrams alongside the answer (via your SearXNG).
• Hybrid retrieval: vector search (sqlite-vec) + keyword search, merged.

Stay organized

• Citation manager — per-subject bibliography with APA/MLA/Harvard formatting and one-click copy.
• Deadlines & calendar — track exams and assignments, with a deadline study checklist and two-way Citations ↔ Calendar sync.
• Tags on topics and deadlines for cross-cutting organization.
• Notes, a lecture recorder with a near-real-time live transcript (~7s adaptive Whisper segments, silence-aware) and an automatic lecture summary saved to Notes after every recording.
• Insights — a study-analytics dashboard: GitHub-style year heatmap of focus hours (pomodoro + passive in-app time), reviews/accuracy/streaks, a 7-day due forecast, and a "topics needing work" ranking.
• Global semantic search (Ctrl+K) across notes, sources, transcripts, events, and materials — exact matches first, vector hits as related content.
• A Pomodoro focus timer and background music: curated study-music stations (lofi, synthwave, jazz, classical, rain, forest, 40 Hz) played through a headless mpv sidecar, plus your own custom stations.

Own your data

• Anki .apkg import and export for flashcard decks, PDF export, and a portable SQLite dump.
• Encrypted homelab backups — snapshot → age encrypt → rclone upload.
• Everything is local; AI is bring-your-own-key; web search is your self-hosted SearXNG.

Built for power users

• Helix-style modal keyboard engine with a command palette, leader keys, and fully customizable bindings.
• 10 themes and live re-skinning (designed to follow the Omarchy palette).
• Close-to-tray — closing the window keeps ingest/generation/music running behind a tray icon (Open · Play/Pause · Quit); reminders become OS notifications while hidden.
• A responsive shell that stays intentional when tiled narrow (drawer sidebar below 1080px).

Download & install

Grab the latest build for your operating system from the Releases page — no toolchain required.

macOS

1. Download Cortex_x.y.z_universal.dmg (works on both Apple Silicon and Intel).
2. Open the .dmg and drag Cortex into Applications.
3. First launch: right-click Cortex → Open (the build isn't notarized yet, so macOS asks you to confirm an unidentified developer once).

Windows

1. Download Cortex_x.y.z_x64-setup.exe.
2. Run it. SmartScreen may warn about an unrecognized app — click More info → Run anyway (the installer isn't code-signed yet).
3. Launch Cortex from the Start menu.

Linux

Distro family	File	Install
Debian / Ubuntu	Cortex_x.y.z_amd64.deb	sudo apt install ./Cortex_*.deb
Fedora / RHEL	Cortex-x.y.z-1.x86_64.rpm	sudo dnf install ./Cortex-*.rpm
Any (portable)	Cortex_x.y.z_amd64.AppImage	chmod +x Cortex_*.AppImage && ./Cortex_*.AppImage

Builds are produced automatically by CI (.github/workflows/release.yml) on every v* tag. They are not yet code-signed/notarized, hence the one-time security prompts above.

Build from source

Prerequisites

• Rust (stable) and the Tauri 2 system dependencies for your OS (on Linux: WebKitGTK 4.1, GTK 3, libsoup3, etc.).
• Node 18+ or Bun for the frontend.

git clone https://github.com/PndaMan/cortex.git
cd cortex
bun install          # or: npm install
bun run tauri dev    # launches the desktop app with hot reload

Build a production bundle:

bun run tauri build

Demo / screenshot mode

Run a throwaway instance pre-filled with rich showcase data (subjects, sources, cheatsheets, flashcards, a full assignment board, and weeks of backdated study activity for the analytics dashboard) — your real library is never touched:

./scripts/demo.sh          # launches against a separate demo database
./scripts/demo.sh reset    # wipe + reseed it from scratch

Under the hood: CORTEX_DATA_DIR points the app at a separate folder and CORTEX_DEMO=1 seeds that empty DB. Set those two env vars yourself to script it.

Most tools ship inside the release builds. yt-dlp (YouTube + music), age, and rclone (encrypted backups) are bundled as sidecars by scripts/fetch-sidecars.mjs — no manual install. The table below applies mainly to running from source or to the few heavier, still-optional tools.

Optional integrations (enable the features you want)

Feature	Needs	Notes
Lecture transcription	local openai-whisper / whisper.cpp / auto faster-whisper, or a homelab Whisper URL	set the Whisper URL in Settings → Integrations to skip local setup
Web search / images	a self-hosted SearXNG with JSON output enabled	set its URL in Settings
Local models	Ollama (local or homelab)	keyless chat + embeddings
Slide previews (optional)	LibreOffice (soffice)	DOCX/PPTX text ingests natively with no tools; LibreOffice is only needed for the faithful rendered-PDF preview. On Windows/macOS you don't have to install it — the doc still ingests, just without slide thumbnails.
PDF text	pdftotext (poppler)	faster & cleaner than OCR
YouTube + music	yt-dlp (bundled) + mpv for playback	yt-dlp ships with the app; mpv still system-installed
Encrypted backups	age + rclone (both bundled)	configure in Settings → Backups

One-command homelab backend. Don't want to set these up by hand? The homelab/ folder has a docker compose stack that runs SearXNG, an OpenAI-compatible Whisper server, and (optionally) Ollama. Bring it up, expose it over a reverse proxy or VPN (Tailscale / Netbird), and point Cortex at it in Settings → Integrations. See homelab/README.md.

Configuration

Open Settings in the app:

1. API keys (bring your own). Paste a key for any of: Google Gemini, OpenRouter, OpenAI, Anthropic Claude, or point at a local Ollama / custom OpenAI-compatible endpoint. Generation works as soon as any key is set.
2. Per-task models. Assign a provider/model to each task (chat, cheatsheet, flashcards, quiz, audio, embeddings) — e.g. a cheap flash model for cards, a stronger one for cheatsheets.
3. Profile. Your name, level, and field personalize tone and examples.
4. Web search. Enter your SearXNG base URL to unlock chat web mode + image fetching.
5. Backups. Add an age recipient public key and an rclone remote to enable encrypted backups.

No keys? The app still runs end-to-end with a clearly-marked offline stub so you can explore the flow.

How it works

src/                       Svelte 5 frontend (runes)
  lib/api.ts               typed Tauri command client (mirrors the Rust commands)
  lib/store.svelte.ts      central app state
  lib/keybinds.svelte.ts   Helix-style modal keyboard engine
  components/  views/       UI

src-tauri/src/
  commands.rs              command surface (ingest, generate, chat, search, export…)
  llm.rs                   LLM providers behind one trait (Gemini/OpenAI-compat/Claude/Ollama, BYOK)
  embed.rs                 embedding providers
  vector.rs                f32 BLOB vectors; sqlite-vec ranks, Rust cosine is the fallback
  ingest.rs                parse → chunk → embed → store pipeline
  repo.rs                  all SQLite access
  anki.rs   backup.rs       .apkg export · encrypted backups
  google.rs  notes.rs       Google API helpers · notes
  calendar.rs  review.rs    deadlines/calendar · FSRS scheduling
  exam.rs  analytics.rs   timed exams + grading · study analytics
  mpv.rs                   background music playback
  migrations/              versioned SQLite schema

• Storage: a single SQLite database in the OS app-data dir, with WAL mode and versioned migrations. Embeddings are stored as little-endian f32 BLOBs — exactly sqlite-vec's compact format — so retrieval ranks in SQL via vec_distance_cosine (with a Rust cosine scan as a transparent fallback).
• Privacy: all content, embeddings, and generated material stay in that local DB. The only outbound calls are to the AI provider you configured and (optionally) your own SearXNG.

Tech stack

Tauri 2 · Rust · SQLite + sqlite-vec · Svelte 5 (runes) · TypeScript · Tailwind CSS · Vite

Roadmap

Shipped (v1.0): the Subjects→Topics→Sources core, the ingestion pipeline, all study-material generators (cheatsheets, flashcards, quizzes, audio, infographics, mind maps, slides), citation-grounded chat with web/image mode, FSRS spaced repetition, Anki import/export, timed graded exams with remark, the Insights analytics dashboard (year focus heatmap, weak-topic ranking), global semantic search, auto lecture summaries, close-to-tray, a citation manager + deadlines with calendar sync, tags, sqlite-vec search, encrypted backups, live lecture transcription, drag-and-drop reordering, and a responsive layout that scales toward mobile.

Next: accounts & sync (CRDT multi-device), a mobile build, and richer calendar/citation workflows.

Contributing

Issues and PRs are welcome. The codebase favors surgical, well-tested changes:

bun run check                       # svelte-check (frontend types)
cd src-tauri && cargo test --lib    # Rust unit tests
cargo check                         # fast type/borrow check

Please keep changes scoped, match the surrounding style, and add a test when you touch backend logic.

License

Apache License 2.0.

---

_{Cortex — study like you own your data, because you do.}