jev-voice-browser · posted Sep 18, 2026 · Jev 1.13 by TypeSafe AI · every demo below is a real run on my machine

Jev AI Voice Browser: Talk To Your Browser

Jev AI Voice Browser: you talk to a real browser, and it moves before you finish your sentence.

You say go to DuckDuckGo, and the page is already loading as the last word leaves your mouth.

You say scroll down a bit, click the models link, open a new tab, and a real browser window does it with your hands off the keyboard.

We set this up on my own machine this morning and ran every test that comes with it, and you're going to see each one on screen.

Later on you'll see the one question that lets it act early without acting wrong.

And stick with me to the end, because you'll see what happens when you tell it to click a Place Order button.

The actual sources ↓
§1Watch it act

The page loads while the last word is still landing

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: a real Chromium window on the left, the live decision panel on the right. “Go to” does nothing. “Go to DuckDuckGo” loads the page. “Search for” waits. The full sentence gets typed into the page's own search box.

§2What it is

Three parts: your voice, a model that decides, and a real browser

From your mouth to a moving browserYour voiceChrome turns itinto words live Jev decidesabout ten questionsevery new wordA real browserPlaywright clicks,types and scrolls
posted on GitHub · Sep 18, 2026by Moritz Krembopen code · MIT licence
§3The model underneath

Jev can't write a single word. It only picks an answer.

Jev, in one lineA situationthe words so far+ what's on the pageA questionwith set answers Jev picks onein under halfa secondHow surea number from0 to 1
Jev 1.13 · TypeSafe AI · out since Sep 15, 2026last week: 15,900 of my keywords sorted for 42 cents
§4What's actually new

It asks again every time a new word arrives

When the decision happensOLD WAY~2 seconds✕ wait until you stop talking✕ send the whole sentence✕ big model thinks it over✕ writes out a plan✕ then the browser moves✕ every command feels lateNEW WAY~0.4 seconds✓ ask on every new word✓ about ten questions at once✓ answers come back as numbers✓ code checks the numbers✓ browser moves mid-sentence✓ stale asks get cancelled
§5One request, about ten questions

Everything it needs to know, asked at the same moment

What Jev gets asked on every new wordWhat?go · search · click · typeWhich thing?one label from the pageWhich site?from a list the code ownsFinished?has the sentence endedTalking to me?or to someone elseRisky?buy · delete · sendHow far?a little · a page · the endWhich words?to type or searchWhich tab?next · previous · first
§6My speed test

Same pages, same eight commands: Jev against three fast models

Middle time to decide, in secondsJev · 9 questions0.45 sClaude Haiku 4.51.6 sGemini 3.8 Flash2.2 sGPT-5.6 Luna2.5 s8 commands · Jev answered 9 questions, the others 1 · Jev 8/8 right, others 7/8
§7The question that lets it act early

“Has this person finished the sentence?”

The “finished?” score, one word at a time“search”0.22“search for”0.03“…for alan”0.91“…alan turing”0.98the app only acts above 0.60 · typing and searching also wait for a short silence
§8On a real page

It uses the search box that's already on the page

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: a Wikipedia article about coffee. “Search for” makes it wait. “Search for espresso” gets typed into the site's own search box, enter is pressed, and the espresso article opens.

§9Searching a site by name

“Search GitHub for Playwright.” The code owns the address.

A named site search“Search GitHub forPlaywright”Jev picks two thingssite: githubwords: “playwright”Code fills the addressgithub.com/search?q=playwright Results openchecked live onmy machine
§10It can't put words in your mouth

The code cuts your sentence into pieces. Jev only points at one.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: “type flat white into the search box”. On the right, under “text to type”, the piece “flat white” scores 1.00, and those exact letters land in the box. Then “press enter”.

How words get into the boxYour sentence“type flat white intothe search box”Code cuts pieces“flat white” ·“white” · none Jev points at one“flat white” · 1.00Copied exactlyletter for letter
§11 Quick one before the use cases

Talk to your tools — and get hours back every week

If you want a browser you can talk to, an inbox that sorts itself and agents that decide in half a second running inside your own business, this is where we build it with you.

The Agent OS zip file — your Claude, your Hermes and your OpenClaw in one dashboard, sharing one memory
My own tools inside it — SEO agents, video, AI avatars — and room to add your own workflows
A video walk-through + 30-day roadmap — for setting the whole thing up
Daily updates — as new versions of the Agent OS land
Four coaching calls a week — bring your screen and we'll wire Jev up with you live
Daily tutorials — covering new setups like this voice browser as they come out
Over 3,000 business owners helped — some had never used AI before they joined
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
258 documented member wins · 38 countries · 4 live calls a week
§12Use case · reading hands-off

“Scroll down a bit.” “Scroll to the bottom.” “Scroll up a page.”

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: three scroll commands in a row. The “scroll amount” bars on the right flip between a little, one page and to the end.

§13Use case · clicking by describing

“Click the models link.” On my own leaderboard.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: goldiebench.com, my AI model leaderboard. “Click the models link” opens Models. “Go back.” Then “click compare”. On the right, the “target element” bars show which label Jev picked.

How it knows what's on the pageSnapshotup to 100 things youcan click or type inShort labelse01 · e02 · e03 … Jev answerswith one labelor “none”
§14When it isn't sure

It doesn't guess. It puts numbers on the page and asks.

Sure, or not sureWhich thing?Jev scores everylabel on the page0.45 or moreclear winner →it clicksUnder 0.45numbered badges onthe top 2 or 3 picks You say “two”code handles it ·no model call
I saw this on a news page with 30 links: “click on a link please” → badges → “the first one”
§15Use case · tabs

“Open a new tab.” “Next tab.” “Close this tab.”

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: four commands, four moves — a new tab, goldiebench.com loads in it, back to the first tab, then the tab closes.

in the full test run, “close this tab” moved after word 2 of 3
§16Use case · my own site, start to finish

“Go to goldiebench dot com.” And the one address it couldn't open.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: the browser goes to goldiebench.com, scrolls down a page, scrolls back to the top and clicks Methodology, all from spoken words.

“go to agentos dot guide” → “where to? no site or domain recognised”the code knows endings like .com .org .ai — .guide isn't on its listthe code owns the addresses · the model only picks
§17Use case · two things in one breath

Eleven words. It started moving at word seven.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: “go to example dot com and click the more information link”. The site loads while the sentence is still arriving, then the link gets clicked when it ends.

When the browser started movingword 7 of 11“go to example dot com and click the more information link”from the full 16-step test run on live sites
§18Use case · it knows when you're not talking to it

“So anyway, I think we should get lunch.” Nothing moves.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: two lines of normal chat get ignored — the “is_command” score sits at 0.01. Then “scroll down” scores 0.99 and the page moves.

“Are they talking to me?”“…we should get lunch”0.01highest, word by word0.06the line it has to pass0.50“scroll down”0.99scored on my machine, one word at a time
§19The Place Order button

“Click place order.” It stops and asks first.

real session on my machine · real Jev calls through OpenRouter · words streamed in one at a time, the way the mic sends them

What you're watching: a local test checkout I made, so nothing real is bought. “Click place order” turns the panel red: CONFIRM. “Cancel” drops it. Asked again, “confirm” clicks it, and the page shows “Order placed (test page)”.

The risky-click gate“Click place order”it finds the buttonRisky? 0.93the line is 0.50 It asks firstsay “confirm”or “cancel”Only thenthe click happens
§20Honest safety notes

The confirm step is a helper. It can still mishear you.

What the builder says, and I'd repeatYour computer onlyit listens on 127.0.0.1Its own profilea separate browser windowConfirm = helpera convenience, his wordsStay signed outof anything that matters
§21“That's for developers”

Four lines, one key, and a Start mic button

What it takes to run4 linesin a terminal1 keypasted into one fileChrome or Edgefor the microphone~1 centfor the full 16-step run
THINKING IT? “I've never opened a terminal. This isn't for me.”

We have people in the Boardroom who had never opened one before they joined, and they run setups bigger than this.

The idea underneath is plain English: write a question, list the answers, and set a line for how sure it has to be.

no TypeSafe key on my machine → ran it on the OpenRouter key I already had, with a 20-line adapter file
§22The numbers from my machine

39 of 39. 27 of 27. 16 of 16.

Every test that ships with it39 / 39quick checks, no internet27 / 27real Jev calls · avg 441 ms16 / 16spoken commands, live sites66 callsto Jev in that live run415 msaverage per call1.2 centsfor the whole live run
§23What a day of talking costs

About two hundredths of a cent per decision

Per decision, measured on my run$0.00019per decision · 66 calls for $0.01231,000 commands × 5 looks each ≈ 5,000 decisions≈ 93 cents a day (my estimate from that rate)
§24Bigger than the browser

The same shape runs an inbox, a keyword list and a task board

One pattern, many screensA browser you talk tothis guideAn inbox that sorts itself200 emails · $0.0041A keyword list colouring in15,900 keywords · 42 centsA board that hands out cards20 cards · $0.0004 The patternsituation in ·set answers ·numbers out ·code decides
The big models still do the writing and the thinking. A deciding model handles the hundreds of tiny calls in front of them.
§25“I'll wait until it's polished”

Sep 15. Sep 18. Sep 20. Nobody waited.

Five daysSep 15Jev comes outSep 18voice browsergoes on GitHub Sep 20running on mymachine · 16 of 16
THINKING IT? “It's a few days old. I'll come back when it's finished.”

The people who learn the pattern this month will know exactly what to build when it shows up inside the tools they already use.

You don't have to run this project to get the lesson. You have to understand the shape.

§26“I don't trust agents to click things”

Every spoken sentence passes five lines you can read

The lines one sentence has to passIs it a command?0.50Is the intent clear?0.55Is it finished?0.60Is the target clear?0.45Risky? then ask first0.50all five are plain numbers in one file · the live panel shows each one for every decision
THINKING IT? “I've had an agent click the wrong thing before.”

Here every decision shows its scores on the panel, next to the line it had to pass.

You can see why it moved, every single time, and you can move the lines yourself.

§27Where it's rough

A very good demo that someone built in a few days

Works today / rough edgesWORKS TODAYon my machine✓ go to · search · click · type✓ scroll · back · reload · tabs✓ two commands in one breath✓ ignores normal chat✓ asks before risky clicks✓ 16 of 16 on live sitesROUGH EDGESbe honest• mic needs Chrome or Edge• first call takes ~700 ms• looks at 100 things per page• knows a short list of sites• tuned to one Jev version• I streamed words, not a live mic
§28Where this goes

Any tool with a screen and a set of buttons

Same deciding layer, different way inYour voicewhat you just watchedA message from another agentsame questions, no micA schedulesame questions, no person The deciding layerunder half a seconda fraction of a cent
§29 Your move

Get your business running on agents you can talk to

You get the Agent OS as a zip file, the video walk-through and the 30-day roadmap — and we'll set up your deciding layer with you on a live call.

The Agent OS zip file — one dashboard for your Claude, your Hermes and your OpenClaw, sharing one memory
Workflows already built in — SEO agents, video, AI avatars — plus your own
The video tutorial + 30-day roadmap — for implementing it step by step
Daily updates — new versions of the Agent OS as we improve it
Four coaching calls a week — bring your Jev setup and share your screen
Daily tutorials — new tools like this voice browser, walked through as they come out
A prompt library + member map — meet people near you building the same things
Over 3,000 business owners helped — some had never used AI before
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
258 documented member wins · 38 countries · 4 live calls a week
THINKING IT? “Doesn't running an agent operating system burn a fortune in tokens?”

No. The everyday work runs on free local models on your own machine, free APIs slot in for more, and the heavy work drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and the Agent OS plugs straight into it.

A deciding model like Jev pushes that lower still — the full 16-step run on this page cost 1.2 cents — and the Boardroom has full token-efficiency tutorials.

§30The close

You talk like a person. The machine keeps up.

What you just watchedActs earlymid-sentenceYour own wordscopied, never inventedIgnores chat0.01 on lunch plansAsks firston buy · delete · send~0.4 sper decision$0.00019per decision
For years we bent ourselves around the computer. This bends the other way.