Locus

The open-source, local-first answer to Google NotebookLM — for video and web sources.

_{开源版 Google NotebookLM · 本地优先 · 视频与网页即来源}

Paste a video link, a web page, or a local media file. Locus keeps it on your machine, transcribes it, and turns it into a searchable notebook with summaries, mindmaps, and grounded insight prompts.

---

Overview

Locus is an open-source alternative to Google NotebookLM. NotebookLM turns the documents you upload into a grounded notebook; Locus does the same job for the sources NotebookLM cannot handle well — video links, web pages, and local media — and does the heavy lifting on your own machine instead of in a cloud pipeline.

一句话：把视频、网页和本地音视频变成你的本地笔记库——下载、转写与文字稿留在本机；摘要、思维导图、要点与解剖由你确认后调用云端 AI 生成。

Note

Latest stable release: v0.6.1 (Windows x64, macOS Intel, macOS Apple Silicon), published 2026-11-11. See GitHub Releases. Licensed under Apache-2.0; the desktop client, worker, and server all live in this repository.

flowchart LR
    A[Video link / web page / local media file] --> B[Download or WebView extraction]
    B --> C[ffprobe and 16 kHz mono audio]
    C --> D[SenseVoice Small local transcription]
    D --> E[Local note]
    E --> F[Confirmed AI steps: summary / mindmap / insights / dissection / draft]

Loading

Important

Locus is for public videos, videos you own, or videos you are authorized to process. It does not bypass platform login walls, CAPTCHA, DRM, copyright restrictions, or privacy boundaries.

Download

Grab the newest installer from GitHub Releases → latest.

Platform	Asset
Windows x64	Locus_0.6.1_x64-setup.exe (plus Locus_0.6.1_x64-setup.exe.sig for the in-app updater)
macOS Apple Silicon	Locus_0.6.1_aarch64.dmg
macOS Intel	Locus_0.6.1_x64.dmg

Releases also carry latest.json for the Windows updater and on-demand pytorch-runtime-*.zip components for the optional PyTorch ASR model. macOS builds are ad-hoc signed, not notarized: the first launch needs a one-time System Settings → Privacy & Security → Open Anyway.

Warning

The account service runs at locus.8xf.pro. Releases up to v0.6.0 have the previous domain compiled in, so they can no longer sign in or use Locus-managed AI steps. v0.6.0 receives the fix through the in-app updater; v0.5.1 and earlier need a manual reinstall. Local state from earlier installs is not migrated: the current data directory is %LOCALAPPDATA%\pro.8xf.app.

Locus vs Google NotebookLM

	Google NotebookLM	Locus
License / hosting	Proprietary, cloud-only	Apache-2.0, an app you install and run
Sources	Uploaded documents, some video URLs	Video links, public web pages, local media files
Where media is processed	Uploaded to Google	On your machine — media, transcripts, model caches stay local
Transcription	Cloud ASR	SenseVoice Small, local
Grounded artifacts	Overviews, mindmaps, reports, study aids	Summary, Mermaid mindmap, insight topics, transcript dissection, platform-style draft
Source-grounded chat	Yes, with citations	Not yet — Locus generates artifacts, not follow-up answers
Multi-source synthesis	Many sources in one notebook	One source per note; collections group notes but do not cross-analyze

Honest boundary: Locus is not a feature-for-feature clone. It covers the source → transcript → notebook → grounded artifacts path and adds local media processing; it does not ship an audio overview, a chat surface, or cross-document reasoning.

What it does

• One home composer for three inputs: a public/authorized video link, a public web page, or one local video/audio file.
• Local notebook with a sidebar (Home / Notes / Collections / Artifacts), pinned notes, search, rename, collections, and export.
• Video download through yt-dlp with focused Douyin / Xiaohongshu / Bilibili fallbacks and a 20-minute per-download budget; valid Xiaohongshu platform subtitles are reused instead of running ASR.
• Web pages loaded in a system WebView and saved as local Markdown with source URL and extraction time.
• Selectable local ASR models (SenseVoiceSmall-ONNX default; PyTorch optional with an on-demand ~140 MB runtime).
• Separately confirmed AI steps — summary, mindmap, insight topics, transcript dissection, platform-style draft — using Locus-managed models with quota checkout by default, or your own OpenAI-compatible provider (BYOK). BYOK keys are stored only in the operating system keychain, and those steps need no sign-in or AI Credit; platform-style drafts still use server-managed search and require a signed-in account. A failed AI step leaves the transcript intact and can be retried alone.
• Automatic LLM note titles, a per-user Web dashboard with email-OTP login, and a local-only diagnostic export.
• Windows installer preflight for free space and writability, with a step log at %TEMP%\install.log.

One source at a time; this is a local notebook, not a general downloader or batch queue. Platform links may still fail when they are private, expired, region-blocked, login-gated, or otherwise unavailable. Full capability and artifact lists live in docs/product-specs/ and docs/ARCHITECTURE.md.

Why Locus

Principle	What it means
Local-first	Videos, audio, transcripts, model caches, and history stay on your machine.
Open source	Apache-2.0; desktop client, worker, and server in one repository; the AI steps can be skipped entirely.
Notebook, not a downloader	Every source becomes a reusable local note instead of a one-off export.
Two-stage AI	Download and transcription run first; each AI step needs a separate confirmation, and Locus-managed steps also check quota.

Quick Start

uv sync --dev
Push-Location app; npm install; Pop-Location
Push-Location server; npm install; Pop-Location
npm --prefix app run tauri -- dev

Before claiming a change is complete, run the full gate set (see docs/EXECUTION_GATES.md):

uv run python scripts/validate_agents_docs.py --level ERROR
uv run python scripts/validate_agents_docs.py --level WARN
uv run ruff check worker
uv run pytest worker/tests
npm --prefix app test
npm --prefix app run lint
npm --prefix app run build
npm --prefix server test
npm --prefix server run build
cargo test --manifest-path app/src-tauri/Cargo.toml
node --test scripts/tests/*.test.mjs
python -m pytest scripts/tests

Build without packaging an installer with npm --prefix app run tauri -- build --no-bundle; the executable lands at app/src-tauri/target/release/app.exe. Local state lives in %LOCALAPPDATA%\pro.8xf.app.

Repository Map

Path	Role
app/	Tauri + React + TypeScript desktop client
worker/worker_core/	Python worker: download, media validation, ASR, and the Studio AI tools (summary, mindmap, insights, dissection, draft)
server/	Fastify account, entitlement, and LLM-checkout service
contracts/	Cross-language task and artifact contracts
deploy/	Production runbook plus Nginx and systemd reference configs
site/	Marketing site (gitignored, deployed separately)
docs/	Product specs, design docs, security, execution plans
AGENTS.md / WORKFLOW.md / TASKS.md	Collaboration map, workflow rules, task ledger

Security Boundary

• Only public, owned, or authorized sources; no login-wall, CAPTCHA, DRM, or privacy bypasses.
• Only explicitly confirmed AI steps may send transcript text to an LLM: Locus-managed steps go through the server-managed provider and quota checkout, while BYOK steps go to the OpenAI-compatible endpoint you configure. The account service never receives media, transcripts, insights, cookies, model caches, or local history.
• API keys for custom providers stay in the operating system keychain; they are never written to config files, logs, command-line arguments, or diagnostic exports.
• Real .env files, SQLite databases, backups, logs, model caches, outputs, and secrets stay out of git.

Full boundary: docs/SECURITY.md.

Documentation

Topic	Where
Product specs	docs/product-specs/
Architecture and design	docs/ARCHITECTURE.md · docs/DESIGN.md
Security	docs/SECURITY.md
Execution plans and hard gates	docs/exec-plans/ · docs/EXECUTION_GATES.md
Server deployment	deploy/README.md · deploy/server-deployment.md
Release workflow	.github/workflows/desktop-release.yml