01
chatai sidebar
a chat panel over any tab that already knows the page — url, title and full text are passed in — plus anything relevant from your brain. streams, renders markdown, and each tab's conversation is saved locally.
built · launching very soon
Aura
an ai that reads the page with you, agents that drive your tabs, and a second brain for everything you save. built on chromium. runs on your machine. your data never has to leave it.
watch
the teaser · 48 s
aura is built. save, connect, ask, act, and keep it yours. launching very soon.
1080p60 · rendered from code · no voiceover
the problem
tabs are not memory. by midnight the forty you opened are gone, and the three that mattered are somewhere in history with the rest.
tuesday09:00
0 tabs open0 remembered
01 / capture
highlight. save. tagged and summarised, on your machine. select text on any page, right-click, and it lands in your brain with the source url attached.
watch
arxiv.org/abs/2401.04088
⌘kMixture-of-Experts, explained
arxiv.org · 12 min read · tue
A mixture-of-experts layer replaces one big feed-forward block with several smaller “experts” and a router. a router picks two of eight experts per token, so the model holds far more parameters than it ever runs for a single token. That is the whole trick: capacity scales with the number of experts, while the cost per token scales with the number of active experts.
Mistral, for example, keeps 47B parameters on disk but runs each token through roughly 13B of them. Training is the hard part: if the router keeps picking the same two experts, the others never learn.
A small load-balancing loss nudges traffic to spread out. Inference has its own problem: all the experts have to live in memory even though most of them sit idle on any given token.
brain
nothing saved from this page yet.
02 / brain
every page you save becomes a node in a graph on your disk. ask in plain words; the answer comes from your pages, with the sources attached. try a question, or hover the graph.
your brain · 46 pages · on disk
|
try another
03 / agent
it browses. it compares. you approve. the agent works on your real tab, shows every step, and stops for a yes before anything with a side effect — a purchase, a form, a new domain.
watch
aura://home
⌘kagent
Bose QC45
$279 · 4.6★
Sony WH-1000XM5
$248 · 4.7★
Sennheiser M4
$299 · 4.5★
Anker Q45
$149 · 4.3★
Beats Studio Pro
$349 · 4.4★
Soundcore Life
$89 · 4.1★
ask, search, or paste a link
agent
04 / private by design
0
pages in your brain
0
bytes sent to the cloud
local ai via ollamaor bring your own api keyseverything stored on your disk
05 / yours
six themes, light and dark, and every workspace can wear a different one — so research looks like research and personal looks like personal. pick one.
aura://home
⌘kHello.
what are we working on?
ask, search, or paste a link
workresearchpersonal
Mixture-of-Experts, explained
1 day ago
Pricing page teardown
yesterday
Standup notes · tue
tue
Headphones under $300
last week
everything inside
02
automationagent mode
describe a task and the agent takes over your active tab: navigates, searches, reads, compares and reports back. you watch every step. anything sensitive pauses for your allow or deny.
03
⌘k · ctrl+kcommand bar
one shortcut anywhere. type what you want in plain language and choose how it runs: drive your live tab step by step, or hand it to a headless browser that streams progress back. failed runs resume.
04
knowledgebrain workspace
write notes, ingest pdfs and logbook entries, and see everything as a knowledge graph of projects, sources and ideas. semantic search over it grounds every ai answer in what you actually saved.
05
isolationworkspaces
work, research, personal — each with its own tabs, theme, quick links and brain. nothing bleeds across. switch from the sidebar and the whole browser changes context with you.
06
readingresearch library
save pdfs and pages into categories and a reading list. open anything in a dedicated reader: the document, ask-ai about it, and your notes, stored next to the source.
07
journallogbook
timestamped entries on a timeline. select text on any page, right-click, and it becomes an entry with the source url attached. a quiet record of what you read and thought.
08
chrome web storeextensions
install straight from the chrome web store by id, or from a .crx file. toolbar popups, enable/disable, one-click update all. the extensions you rely on come with you.
09
modelsbring your own model
local models run through ollama and cost nothing — on the free plan too. add an openai or anthropic key and gpt-4o or claude appear in the same dropdown, used only when you pick them.
10
browserthe basics, done properly
isolated chromium tabs, session restore, reopen closed tabs, idle-tab suspension, find in page, downloads, bookmarks, history, quick links, shortcuts for everything, background auto-updates.
how it works
shell
electron + chromium
the ui is react; every tab is a real, isolated chromium process. frameless window, native feel on windows and macos.
sidecar
local ai runtime
a python service on localhost runs chat, the agent loop (playwright), the brain index and vector memory. it starts with the app and talks to nothing but your machine.
storage
your disk
history, bookmarks, notes, pdfs, chat logs, embeddings and models live in plain files and sqlite under your user folder, per workspace. export, back up or delete at will.
one small cloud service — the aura engine — exists only to hold your account email and plan so a subscription works across the desktop app and this site. it never sees a url you visit.
what leaves your machine
datagoes to
browsing history, bookmarks, tabsyour machine only
notes, logbook, library pdfs, brain graphyour machine only
page text you chat aboutlocal model by default · cloud only if you pick a cloud model
api keys for openai / anthropicstored locally, sent only to that provider
account email and planaura engine (needed for sign-in and billing)
update checkversion number only
there is no analytics in the browser. clear history, cache and cookies from settings at any time, and every store is a file you can open yourself. read the privacy policy.
under the hood
- engine
- chromium via electron
- ui
- react 19 · tailwind · framer motion
- local inference
- ollama (any installed model)
- cloud inference
- openai · anthropic · your own keys, optional
- agent runtime
- playwright + drive-tab on your live tab
- memory
- sqlite index · vector embeddings · on disk
- extensions
- chrome web store compatible
- telemetry
- no browsing data · ever
- updates
- signed feed · sha-512 verified · background install
- platforms
- windows · macos · linux
Aura
Aura is built. Leave your email and you'll be among the first to get it on launch day.
launch news only · no spam
aura.thebuilder.company
windowsmacoslinux