ask about a screen you already closed.
it kept a thumbnail of what mattered. the tab doesn’t need to still be open.
ask anythingabout what you see.
you keep alt-tabbing to chatgptjust to describe what’s on your screen. harness already sees it. point it at any screen, window, or tab and ask.
no install · no card · any browser
any screen, any window, any tab. your call.
examples of what it can be pointed at. not endorsements or partnerships.
runs where you are
it runs on your machine.
other screen assistants are native apps that demand deep system permissions to watch you. harness is just a browser tab. the watching, the remembering, and the deciding-what-matters all run on your own hardware. when you ask, only the few frames that answer your question leave, anonymized.
your compute no app. it runs in the browser.
screen capture, the AI that reads your screen, the keyframe memory, the search over it: all of it runs in the tab you already have open. you grant a screen-share once, at the browser level, and that's the whole permission story. nothing to install, nothing running in the background with os-level access.
one browser share, no os permissions
your screen it doesn't stream your screen.
harness watches locally and keeps the continuous capture on your device. when you ask something, it picks out just the few frames that actually answer your question and sends those, with your prompt, to venice. nothing else goes, and the request is private so it can't be tied back to you.
only the frames that matter, to venice
your memory it remembers you, on your device.
what harness learns about you lives in a database on your own machine, not on our servers. want a backup? it's encrypted with a key only you hold, so all we ever store is a blob we can't read.
running on-device today
what it sees
point it at anything.
it doesn’t matter what’s on the screen, a chart, a contract, a game, homework. it reads what you’re looking at and answers. six real conversations on six real screens, below.
- 01
the trading desk
“what do you think, long or short?”
- 02
a prediction market
“see any mis-priced markets?”
- 03
your game in progress
“what's a strong developing move?”
- 04
studying for the gmat
“walk me through these questions.”
- 05
stuck on a calc problem
“help me solve these step by step.”
- 06
shopping the odds
“any value on tonight's game?”
the ask
ask anything about what you’re looking at.
it already sees what’s on your screen, so you can skip the describing and just ask.
- 01
“what changed on this chart since i was last looking?”
hyperliquid · BTC-PERP · 15m
- 02
“what does this clause actually require of us?”
vendor-msa.pdf · §12 indemnification
- 03
“is this market mispriced or am i high?”
polymarket · 2026 election market
- 04
“why is this throwing this error?”
vscode · stack trace
screen memory
it remembers what it saw.
harness keeps a short, searchable memory of your screen. reach back to something from ten minutes ago, or hand it a stretch of time and let it watch.
hand it a stretch of time.
tell it what to watch for and what you want back, a heads up, notes, a summary. it keeps an eye out and reports. it never clicks or types, it only watches and tells you.
vs. the old way
you’re already doing this.
there’s a faster way.
the same question, asked two ways. one takes eight steps. the other takes two.
01 alt-tab away from the chart
02 ⌘-⇧-4 to drag a screenshot
03 open chatgpt.com in a new tab
04 drag the screenshot into the input
05type “what is this chart showing?”
06 wait for the upload
07read an answer that doesn’t know your context
08 alt-tab back to the chart
01ask “what changed?”
02 read the answer that’s about your chart
the chart is still in front of you.
you never alt-tabbed.
memory
it gets to know you.
most assistants forget you the moment you close the tab. harness doesn’t.
memory it remembers you.
harness keeps a private memory of what you were working on, what you asked, what you liked, stored on your own device. open it tomorrow and it picks up where you left off.
“we were halfway through chapter three last night. ready to pick up?”
profile it learns who you are.
tell it once that you trade on hyperliquid, that you’re studying for the bar, that you prefer terse answers. it carries that context everywhere you take it.
“right, you don’t care about the index funds. here’s the spicy stuff.”
skillssoon it can act on what it sees.
give it skills and it does more than answer: research a question across the web, pull the data, draft the message, kick off a workflow. it picks the right tool and runs it, and it still never takes over your machine.
“pulled the three comps and drafted the recap for you.”
where your screen goes
watched locally.
answered privately.
what stays on your device, what leaves, and who can see what. three steps.
you point harness at one thing.
it only sees the screen, window, or tab you choose. not your other tabs. not your desktop. nothing in the background.
the looking happens on your device.
harness runs its own AI inside your browser to watch, remember, and pick out what matters. when you ask a question, what leaves is your prompt and a focused slice of the relevant part of your screen, nothing more. that goes through Venice, a private inference provider whose infrastructure is anonymized by default, so the request can't be tied back to you.
the memory never leaves your device.
your memory stays on your machine. what harness learns about you (your preferences, your recurring contexts, the thread of your work) lives in a database on your own device, not on our servers. turn on backup and it's encrypted with a key only you hold, so all we ever store is a blob we can't read. wipe it whenever you want.
what people say
early users, in their own words.
i've stopped describing charts to chatgpt. it sees what i see. i ask. it answers. about a week in and it already feels broken to go back.
the memory thing is what got me. it remembered i was studying for the bar and started skipping the basics on its own. no other AI does that.
it sits over my editor while i work and only speaks when i ask. zero ceremony. and only what i ask about ever leaves my machine.
questions
the questions you’d ask before trusting it.
01is harness actually private, or is that marketing?
your full screen never streams up, and your memory never leaves your device. the AI that decides what matters runs in your browser, on your machine, and trims what goes out. when you ask a question, what travels is your prompt plus a focused slice of the relevant part of your screen, nothing more. that request goes through Venice, a private inference provider whose infrastructure is anonymized by default, so the call can’t be tied back to you. and soon the client is open source, so you won’t have to take our word for any of this.
02which AI is doing the answering?
two layers. the watching layer is small vision models running in your browser (a CLIP encoder plus a compact captioner), they read your screen locally and never leave your machine. the answering layer is your call: run a model locally on your own hardware, bring a venice or bankr key, or use managed credit and we route to frontier models (gemini, claude, gpt, grok) through Venice’s anonymized infrastructure, no key needed.
03what does “bring your own key” mean?
harness reaches AI through Venice and Bankr, the routers that cover both open models and the closed frontier ones. if you have a venice or bankr key, plug it in and harness spends your own credit instead of ours. the open tier is free this way, and fully free if you run a model locally on your own machine. prefer not to bring a key? use managed credit instead, pay-as-you-go, starting with $5 free. resident adds encrypted cloud backup you can restore on any device.
04can it look back at a screen from earlier, or watch something for me?
yes, both. harness keeps a short, searchable visual memory of what it has seen, so you can ask about a screen from ten minutes ago without having it open again. you can also point it at something and say watch this for twenty minutes, note when the number crosses a threshold, then write me a summary. it keeps an eye on it and reports back. it never clicks or types for you, it watches and tells you what it saw.
05do i need to install anything?
no. harness runs in any modern chromium browser (chrome, brave, arc, edge). open the tab, share what you want it to see, ask. that’s the whole install.
06what does it remember about me?
your preferences (how you like answers shaped), recurring contexts (you trade on hyperliquid, you study for the bar), and the thread of your work: what you asked yesterday, what you were on. not your screen, not screenshots, not your typing. just the gist, stored in a database on your own device, wipe-able at any time.
07is it open source? can i run it myself?
the harness client is being open-sourced. you’ll be able to read every line that watches your screen and holds your memory, and run the whole thing on your own infrastructure if you’d rather not use the hosted version. the hosted product and the self-hosted one are the same client. that’s the point: you don’t have to trust us, you can check.
08what if i cancel?
you drop back to the open tier and keep everything that runs on your device: your local memory, your own model or key, the floating window. what pauses is the paid layer, the cloud backup, or managed AI. your memory is local, so it’s never held hostage. for a full delete, one click in settings.
pricing
start free. pay only if you want more.
everything that makes harness work is free and runs on your device. managed AI is pay-as-you-go credit, and the only thing you pay monthly for is backing your memory up and carrying it across devices.
open
your AI, your machine. free forever.
Free
- run your own model locally, or bring a venice or bankr key
- on-device memory that never leaves your machine
- $5 of managed inference to start, top up anytime
- floating window over any screen, window, or tab
- most picked
resident
your memory, everywhere.
$0.00/ month
- everything in open
- encrypted cloud backup, with a key only you hold
- pick up where you left off on any device
- $5 of managed inference included every month
pay as you go
managed AI, only what you use.
from $10
- top up managed inference in $10, $20, $50, or $100 packs
- credit never expires, no subscription
- the model's list rate, nothing on top
- works on every plan, free or resident
one click away
see what your screen could tell you.
no install. point it at any screen, window, or tab you choose.
or read the code and run it yourself. open source, soon.