noderesearch.app · Download · Free and open source for macOS, Windows and Linux
NODE Research Browser (NODE for short) is a desktop browser for knowledge networking. It maps what you read on a canvas, and you send the whole trail to a friend or a small group as one file. They open it in NODE, replay how you built it, and save their own copy to keep going. NODE also reads the page you are on, picks out its keywords, and suggests the next pages to read.
Built with Electron and plain JavaScript, with an Expo mobile app (iPhone and Android, in testing) that shares the same core logic. Status: released, version 0.10.0, updated often.
| Suggestions from the page you're reading. Keywords come from the part of the page in view. Pin a keyword to keep it, × to hide it, or select text to search for it. Results come from Brave Search, Wikipedia, SearXNG or DuckDuckGo, with image previews. |  |
| Mosaic: four tabs playing at once. Tile tabs in a grid. Media starts on its own and up to four tabs play together; "Sound: all" lets you hear every one of them. Clicking a video scrolls it to the middle. |  |
| Research trail. Every page read in a session, with its top keywords, the keywords that recur across pages, and an export to Markdown. |  |
Also: everyday browser basics (right-click menu, downloads, reopening last session's tabs, sleeping idle tabs, print and PDF), a read-only MCP server so AI agents such as Claude Code can search, read pages and read your research trails (details), per-site behaviour rules, ad and tracker blocking, bookmarks, picture in picture, trackpad and mouse gestures, Ask AI about a page with a local model, and a separate Google window for Google sign-in. The full list is in docs/07-features.md.
The interface follows the NODE brand guideline (black ground, Barlow and JetBrains Mono, one violet accent). The screenshots use a fictional site made for them (docs/media/src), so they can be rebuilt with node docs/media/src/make-shots.js.
- Overview, roadmap and backlog, architecture
- Review findings, change log, mobile plan
- Features and competitor comparison, engine decision: Electron or Chromium, releasing
NODE was called Scout until October 2026. The bundle ID app.scout.browser and some internal names keep the old name.
Download NODE for your computer at noderesearch.app, or pick a file from the latest release: .dmg for macOS (Apple chip or Intel), .exe for Windows, .AppImage or .deb for Linux. The Mac app is signed and notarized by Apple. NODE updates itself on macOS and Windows. How releases are made: docs/09-releasing.md.
If you have used NODE for a few days, please fill in the feedback survey (about 4 minutes). It decides what gets fixed and built next. No GitHub account, or prefer to answer privately? Use the anonymous survey instead.
Found a bug? Use Help → Report a Problem in NODE, or open a bug report.
This is a desktop browser (Electron), not a website.
cd scout # the folder that holds package.json
npm install
npx electron .npm start also works after npm install. The first npm install downloads Electron into node_modules — without that step you get electron: command not found.
On macOS the first launch may be blocked by Gatekeeper. If that happens:
xattr -cr node_modules/electron
npx electron .Requires Node 18+.
Shot (or press S anywhere outside a text field) captures the active tab into a real folder on this computer and stores the page URL. Library → Open source goes back to that page. Show folder opens the directory in Finder or Explorer. Save writes a JSON note of the page. Media downloads the largest video or image currently in view (streamed video such as YouTube cannot be saved as one file).
| Keys | Action |
|---|---|
| Cmd/Ctrl+T | New tab |
| Cmd/Ctrl+W | Close tab |
| Cmd/Ctrl+L | Focus the address bar (suggests pages from history as you type) |
| Cmd/Ctrl+R | Reload |
| Cmd/Ctrl+[ and ] | Back, forward |
| Cmd/Ctrl+F, Cmd/Ctrl+G, Shift+Cmd/Ctrl+G | Find in page, next match, previous match |
| Cmd/Ctrl+= , Cmd/Ctrl+-, Cmd/Ctrl+0 | Zoom in, out, reset (pinch and Ctrl+scroll also work) |
| Cmd/Ctrl+Y | History |
| Cmd/Ctrl+Shift+M | Mosaic on or off |
| Cmd/Ctrl+Shift+A | Sound from all playing tabs, or only the current one |
| Cmd/Ctrl+D | Bookmark the page |
| Cmd/Ctrl+Shift+B | Bookmarks |
| Cmd/Ctrl+Shift+P | Picture-in-picture |
| Cmd/Ctrl+Shift+U | Mute or unmute the tab |
| Cmd/Ctrl+Shift+S, or S | Screenshot the tab |
| Cmd/Ctrl+Shift+. | Show or hide the suggestions widget |
| Cmd/Ctrl+, | Menu (opens at Settings) |
| Ctrl+Cmd+F (Mac), F11 (Windows, Linux) | Full screen: the toolbar hides; move the pointer to the top edge to show it |
| Folder | What it is |
|---|---|
core/ |
Logic shared by desktop and mobile (keywords, search sources, AI, rules, sessions, history, page scripts). See core/README.md |
renderer/, main.js, preload.js, view-preload.js, fs-storage.js |
Desktop app (Electron) |
mobile/ |
Mobile app (React Native, Expo) walking skeleton. See mobile/README.md |
test/ |
Unit demos and desktop end-to-end tests |
docs/ |
Overview, roadmap, architecture, review findings, change log, mobile plan |
The mobile app is planned with React Native (Expo) and reuses core/. A first version is in mobile/: try it on your phone with the Expo Go app (cd mobile && npm install && npx expo start, then scan the QR code). The plan is in docs/06-mobile-plan.md.
The old Tauri setup (src-tauri/, renderer/tauri-bridge.js) does not compile and will be removed.
- Toolbar buttons are icons. Hover over one to see its name and shortcut. From the left: back, forward, reload, the address bar, bookmark star, then Mosaic, Sound, Capture (screenshot, save page, save media, picture-in-picture in one menu), Agents (only while agents are on), Downloads (only while one is running), the suggestions widget, full screen and the menu.
- Suggestions widget. Suggest, Ask, Agent and Talk are tabs at the top of the widget beside the page. The arrow at its top right collapses it and gives the page the full width. While it is collapsed, the widget button in the toolbar shows the number of results, and in violet how many arrived since you collapsed it.
- Menu. The menu button (☰) opens Library, Bookmarks, History, Downloads, Trail, Settings and Help in one panel over the page. Esc, × or a click outside closes it.
- Full screen. The toolbar hides. Move the pointer to the top edge to bring it back; it hides again when the pointer moves down. It stays while you type in the address bar or the find bar, or while a menu is open. A video's own full screen covers the whole window.
- Each tab is an Electron
WebContentsView, a real Chromium page. Tabs that are not on screen are detached from the window, so Chromium stops drawing them and slows their timers; their media keeps playing. In Mosaic every tile is on screen, so several videos play together. - Mosaic tiles every open tab. Click a tile to make it the current tab. Focus shows one tab at full size.
- Playing in several tabs: media autoplays in every tab, and up to 4 tabs play at once. When a fifth tab starts, the tab that started playing longest ago is paused and marked ⏸ in the tab bar (the active tab is never paused this way). Opening a paused tab resumes it.
- Sound: by default only the active tab is audible and other tabs keep playing muted. The Sound button in the toolbar (or Cmd/Ctrl+Shift+A) makes every playing tab audible at once. The speaker button on a tab overrides this; right-click it to go back to automatic.
- Autoplay is allowed without a prior click. As soon as a page is parsed, NODE plays its main media with sound: the largest visible video or YouTube, Vimeo or Dailymotion embed, or else the first audio player. Players that appear later (built by a script, or on sites like YouTube that change address without reloading) are caught for 20 seconds. Only the main player starts, and media you pause stays paused.
- Center on click: clicking a video, an image or an embedded player (YouTube, Vimeo and similar) smoothly scrolls the page so it sits in the middle of the view. Small images such as icons are ignored. Site rule
centerturns it off per site. - Popups open as a new tab only if you clicked or typed in the page in the previous 2 seconds. Other popups are blocked and listed in the sidebar with an "Open in new tab" button.
- Permissions: camera, microphone, location, notifications, clipboard reading and screen capture always ask first. Answers can be remembered per site (File → Reset Site Permissions clears them).
- Keywords (
keywords.js): NODE reads the page title, description, headings, selected text and the text currently on screen (navigation and footers skipped). Words are scored by where they appear, how specific they are, and how rare they are across the pages you have visited this session. Multi-word names ("Marie Curie") are kept together. - Pin and hide: click a keyword to pin it; pinned keywords go into every search and are never replaced. × hides a keyword for good (Show all brings them back).
- Search: pins and the top keywords become 2 or 3 short queries (at most 4 words each). Wikipedia runs for every query. Web results come from the first working source in this order: Brave Search API (the default; needs your API key), SearXNG, DuckDuckGo. Searches typed in the address bar open Brave Search (search.brave.com). A source that rejects a key, hits a rate limit or starts blocking is paused for a while and the next one is used. Results are cached for 30 minutes and requests are spaced out per site.
- Results show Wikipedia first, within about a second; web results are merged in when they arrive.
- Previews on result cards load only when a card scrolls into view, 3 at a time. Wikipedia and Brave results that already include an image skip the fetch.
- Auto re-searches when you open a new page, select text, or scroll to content with different keywords (at most every 20 seconds). Turn Auto off to search only when you press Crawl.
NODE has no account. The only usage data it sends is one anonymous count a day (below), which you can turn off in Settings → Help. The usage summary in Help stays on your computer until you copy it. Diagnostics are only copied when you click Copy Diagnostics.
What leaves your computer, and when:
| What | Sent to | When |
|---|---|---|
| Search words: your pinned keywords and the page's top keywords (never the page text) | Wikipedia, DuckDuckGo, Brave Search (with your key), your SearXNG server if you add one | While suggestions are on. A site rule can turn suggestions off for a site |
| A request for each suggested page, to find its picture | The suggested pages' sites | When suggestions show. No cookies, no sign-ins |
| Pinned keywords, up to 8 page keywords and the page title | Your local model, or xAI (Grok) with your key | Only if AI query ideas are on (off by default) |
| The page's readable text, title and address | Your local model, or xAI (Grok) | Only when you ask a question in Ask AI or run the research agent |
| The text, titles and addresses of the tabs in Mosaic, or of the 8 pages you read most recently in the research session | Your local model, or xAI (Grok) | Only when you ask about Mosaic tabs or the session in Ask AI. Session pages that are not open are read again, logged out. Sign-in, account and payment pages are never sent |
| The address of the page you're reading | Hacker News search (Algolia) and Bluesky search | Only while the Talk tab is open |
| A check for a new version | GitHub | Every 6 hours in installed copies, if "Check for updates" is on |
| One anonymous count: NODE's version, your operating system (Mac, Windows or Linux), and whether it is the first count this week and this month. No id, pages, addresses or keywords | NODE's share server (a Cloudflare Worker), which keeps a count per day and nothing else | Once a day while you use NODE, in installed copies. On by default; turn it off in Settings → Help ("Send one anonymous count a day") |
| A download of the ad and tracker block lists (uBlock Origin's default lists; with cookie notices on, also the cookie-notice lists) | GitHub | About once a week |
| Google sign-in and Google apps | A separate Chrome-based browser on your computer, with its own NODE profile | When you open a Google account page (Settings → Google window) |
| A shared trail, encrypted on your computer (the key stays in the link, so the server cannot read it) | NODE's share server (a Cloudflare Worker), which deletes it after 30 days | Only when you click Create Link in Share Trail. Stop Sharing deletes it sooner |
| The id of a shared trail | NODE's share server | When you open a share link someone sent you |
Saved passwords are never sent anywhere. They are encrypted with your computer's own secure storage (Keychain on a Mac, Windows' data protection, a keyring on Linux) and kept on this computer. A saved login is filled only on the site it was saved for, and only when you pick it.
AI agents on your computer can use NODE through its MCP server only if you turn it on. It listens on your computer only (127.0.0.1), never on the network, and agents read pages logged out.
Ads and trackers are blocked on every page with the filter lists uBlock Origin turns on by default, and leftover ad boxes are hidden. The shield next to the address shows how many were blocked on the page; click it for the sites blocked and the switch.
- Bookmarks. ☆ next to the address bar (or Cmd/Ctrl+D) bookmarks the page; ★ means it is saved. Bookmarks appear first in address-bar suggestions and in the Bookmarks shelf of the Library (Cmd/Ctrl+Shift+B).
- Picture-in-picture. Capture → Picture in Picture in the toolbar (or Cmd/Ctrl+Shift+P) puts the page's video, or an embedded YouTube or Vimeo player, in a floating window. Again to close it.
- Ad and tracker blocking. On by default, using the filter lists uBlock Origin turns on by default (uBlock filters, badware, privacy, quick fixes, unbreak, EasyList, EasyPrivacy, Peter Lowe's list) through Ghostery's filter engine, downloaded on first run and updated weekly. Ad and tracker requests are blocked, and leftover ad boxes are hidden on the page. Settings → Also hide cookie notices adds the cookie-notice lists. Turn it off in Settings, or for one site with a site rule
"adblock": false. Settings shows how many requests were blocked. - Passwords. After you sign in, NODE asks whether to save the password (Save Password, Never for This Site, Not Now), and asks to update it when you sign in with a new one. Next time, click the username or password field and pick the login from the menu, or right-click the field → Fill Saved Login. Sign-ins in two steps (username first, then password) and sign-in boxes inside a page from another site work too. On a sign-up or change-password form, click the password field and choose Use Strong Password. Settings → Passwords lists the saved logins with Show, Copy and Remove (Show and Copy ask for Touch ID, or a confirmation on computers without it; a copied password is cleared from the clipboard after a minute), and the sites you chose Never for. Import Passwords… reads a passwords file exported from Chrome, Brave, Edge, Safari, Firefox, 1Password or Bitwarden, and offers to move the file to the Trash afterwards. Without secure storage on the computer, NODE does not save passwords.
- Ask AI across tabs. The Ask tab can also answer about all the tabs in Mosaic or the pages of your research session. Answers cite their sources as [1], [2] and so on; click a number to open that page. Session → Summarize session writes a summary with key findings, disagreements and open questions, cited, and saves it to the Library.
- Ask AI. The Ask tab in the suggestions widget opens a chat about the current page, with Summarize, Key facts and What next? buttons. It uses the model set up under Settings → AI (a local model or Grok). Questions send the page's readable text, title and address to that model; nothing is sent until you ask.
Build an installer on the computer it is for:
npm install
npm run dist:mac # macOS: dist/NODE-0.4.0-arm64.dmg (or x64 on Intel Macs)
npm run dist:win # Windows: dist/NODE Setup 0.4.0.exeThe app is not signed with an Apple Developer ID yet, so the first time macOS says it cannot check the app: right-click NODE in Applications, choose Open, then Open. Automatic updates need a signed app and a download host; they come later.
Desktop:
| Gesture | Action |
|---|---|
| Two-finger swipe right / left on a trackpad | Back / forward. An arrow at the edge of the page fills as you swipe; let go when it turns blue. Not triggered inside areas that scroll sideways |
| Three-finger swipe (macOS, if set in System Settings) | Back / forward |
| Mouse side buttons | Back / forward |
| Pinch, or Cmd/Ctrl + scroll | Zoom |
| Middle-click or Cmd/Ctrl+click a link | Open it in a new background tab |
| Middle-click a tab | Close it |
| Double-click the empty part of the tab bar | New tab |
| Drag a tab | Reorder tabs (also the Mosaic order) |
| Click a video, image or embedded player | Scroll it to the middle of the view |
Mobile:
| Gesture | Action |
|---|---|
| Swipe left / right on the address bar | Next / previous tab |
| Swipe from the screen edge (iPhone) | Back / forward |
| Android back button | Back |
| Pull down at the top of a page (iPhone) | Reload |
| Press and hold the tab count | New tab |
| Swipe a tab left in the tab list | Close it |
| Pinch | Zoom the page |
| Tap a video, image or embedded player | Scroll it to the middle |
Trail in the menu shows the current research session: every page you read (with its top keywords), keywords that recur across pages, and how many suggestions you opened or hid. Click a keyword to pin it. Pinned keywords are saved with the session.
On result cards, × hides a result for the rest of the session, and results you have already opened or read are dimmed and marked "read". New starts a fresh session; the dropdown switches back to an older one. Export saves the session as Markdown notes in the Library (Saves).
The Talk tab in the suggestions widget shows existing discussion of the page you are reading: Hacker News stories and Bluesky posts that link to it, plus a link to search Reddit. NODE looks the page up only while Talk is open, sends its address without tracking parameters, and never looks up sign-in, account, payment or banking pages or addresses on your own network. Turn it off for a site with the site rule "talk": false. This is stage 1 of the social layer plan; posting and accounts come later.
The Agent tab in the suggestions widget asks for a question, then runs a few searches with your sources, reads the best results in a separate logged-out session, searches once more if its notes leave gaps (8 pages at most) and writes a short report with numbered sources. A small live view shows the page it is reading, and you can ask follow-up questions that build on the sources it already read. You see each step as it happens and can stop it. The report is saved to the Library and the research session, and Open sources in Mosaic opens the sources side by side. It uses the model set in Settings → AI.
It is tested against 12 prompt-injection pages before each release (npm run test:redteam), including with a model that obeys them. It only reads pages that came back from its own searches: the model picks results by number and never supplies an address, so hidden text on a page cannot send it somewhere else. It does not sign in, click or fill forms, and skips pages that ask for a password or card details.
NODE can serve AI agents on your computer, such as Claude Code, Codex or Cursor, through the Model Context Protocol. It is off by default and read-only.
-
Settings → Agent access (MCP) → tick Let AI agents on this computer use NODE.
-
Copy the command shown and run it in Terminal. For Claude Code it looks like:
claude mcp add --transport http node http://127.0.0.1:47615/mcp --header "Authorization: Bearer <token>"Other clients take the same address and header in their MCP settings.
-
To let an agent read a page you have open, click the Share tab button in the toolbar (the arrow out of a box). Sharing ends when the tab goes to another site.
What agents can do:
| Tool | What it does |
|---|---|
search_web |
Searches with your NODE sources (Brave, Wikipedia, SearXNG, DuckDuckGo) |
fetch_page |
Reads a public page as Markdown in a separate, logged-out session that never sees your logins |
list_shared_tabs, read_tab |
Lists and reads only the tabs you shared |
list_research_sessions, get_research_session |
Your research trails as Markdown (also available as resources node://sessions/...) |
Limits enforced by NODE, not by the agent: no clicking, typing or signing in; sign-in, account, payment and banking pages and pages with password or card fields are never read; addresses on your computer or local network are refused; only programs with the token can connect, and web pages cannot; calls are rate-limited. Every call is listed in Settings and saved in the research session. The green Agents button in the toolbar (or View → Stop All Agents) turns the server off and takes back every shared tab.
Settings → Site rules lists the rules in plain words. Rule for this site starts one for the page you are on; every setting can be left as "Same as before" so earlier rules still apply. Rules are saved to rules.json, which you can still edit by hand (Edit rules file).
NODE reads rules.json (File → Edit Site Rules…, or Settings → Edit rules file) and reloads it when you save. Each rule has a match and any of these settings:
| Setting | Values | Effect |
|---|---|---|
autoplay |
on, muted, off |
Start media on load, start it muted, or pause media the page starts before you interact |
sound |
auto, on, off |
Default sound for tabs on this site (the speaker button still overrides it) |
popups |
gesture, block, allow |
Popups only after a click, never, or always |
suggest |
true, false |
false stops NODE reading the page for keywords or sending its words to search |
mosaic |
true, false |
false keeps the site out of the Mosaic grid unless it is the active tab |
center |
true, false |
true (default): clicking a video, image or embedded player scrolls it to the middle of the view. false turns this off for the site |
adblock |
true, false |
true (default): block ads and trackers. false allows them on the site |
extract.include, extract.skip |
lists of CSS selectors | Extra text to read with priority; areas to ignore |
match can be example.com (with subdomains), *.example.com (subdomains only), example.com/path, a list of these, * for every site, or a regular expression like /\/login/i. Rules apply in order and later matches win. The first run creates examples for Wikipedia, YouTube, and sign-in, banking and payment pages (no suggestions, no autoplay). A broken file is reported in the sidebar and the previous rules stay in use.
Optional (Settings → AI query ideas). An AI model reads your pinned keywords, the page's top keywords and its title, and suggests up to 2 extra search queries that look for other angles: background, related people or organisations, primary sources, opposing views. The status line marks these queries "(AI)". Keyword-only results still show first, so the AI never slows down the first results.
| Option | Setup | Privacy |
|---|---|---|
| Local model | Install Ollama and run ollama pull llama3.2 (or any chat model). Address http://localhost:11434/v1. LM Studio also works at http://localhost:1234/v1. |
Nothing leaves your computer |
| Grok (xAI) | API key from xAI. Press Test to see the model names your key can use and pick one. | Keywords and page titles are sent to xAI |
The model never receives page text or addresses, and pages with suggestions turned off by a site rule are never sent. If the model is unreachable or answers badly, NODE uses its normal keyword queries and shows a note. The Grok key is stored like the search keys (encrypted with the system keychain when available).
Open Menu → Settings to turn sources on or off. Test runs one search with that source and shows the result or the error.
| Source | Setup | Notes |
|---|---|---|
| Wikipedia | Nothing. Optional language code (en, es, fr…) | Always runs; gives the reference results |
| Brave Search API | API key from the Brave Search API dashboard (api-dashboard.search.brave.com) | Default web source, on from the start. Best web results; has a free tier with a monthly limit. Without a key, DuckDuckGo is used |
| SearXNG | URL of an instance you run or trust, with the json format enabled in its settings.yml |
No key; you control the engines |
| DuckDuckGo | Nothing | Fallback only. Reads the HTML results page, so it can break or be rate-limited |
API keys are stored in NODE's settings.json, encrypted with the system keychain when one is available.
It is not a general-purpose crawler, scraper farm, or archive bot. It only suggests a short list of next URLs from a keyword query against public search. It does not log in, bypass paywalls, or walk an entire site unattended.