Skip to content
Log in

New

How detectable are you right now?

Browser infrastructure for AI agents

Unlimited profiles. Unlimited possibilities. Everything you can do in a browser, your agent can do too. Way faster and at scale.

See the measurements
Download and run one real task
Over MCP, an SDK or direct CDP
Download and run a task like this

  • MCP

    Point Claude, Cursor or your own agent at it, over HTTP or stdio.

  • SDKs

    Python and Node, both on direct CDP, plus a plain REST API. Keep your agent logic, swap the browser underneath it.

  • Direct CDP

    Already running browser-use? One line of connection code and the run starts signed in.

0

Credentials or cookies that leave your machine

11x

Less spent on model tokens for the same hundred steps

18

Captcha types solved on live widgets, not test keys

8/8

Detector sites passed in our own audit bench

240 teams run their agents on it

Stripe
Notion
Salesforce
Datadog
Google
Spotify
Shopify
Uber

#01 - The browser

Give every agent a machine that stays the same one.

The identity comes from the profile seed. Same seed, byte-identical values across restarts and across machines.

seed b0fb2d81-4ec2-41bf-a865-5acf4ae947c8

Persona

Coherent fingerprint axes, derived from the seed. Read-only.

windows
Runs on
Local native Windows
Fixed at creation: moving a profile to another tier means creating a new one.
User agent
Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/148.0.0.0 Safari/537.36
Platform
Win32
Chrome
148.0.7778.215
Screen
1920×1080 @2x
CPU / Memory
8 cores · 8 GB
Timezone
Europe/Vienna
Locale
de-AT
WebGL renderer
ANGLE (NVIDIA, NVIDIA GeForce GTX 1650 (0x00001F0A) Direct3D11 vs_5_0 ps_5_0, D3D11)
GPU
NVIDIA NVIDIA GeForce GTX 1650 (d3d11)
WebGPU
NVIDIA GeForce GTX 1650
Fonts / Voices
58 fonts · 19 voices
Noise seed
17265654078529133351

One browser per profile: its own logins, its own cookies, its own device. Two of them share nothing.

The GPU and the screen are the ones in your machine. A profile that contradicts itself refuses to start.

Engine updates come with the plan. Pin a version, verify it, roll back, and only the difference gets downloaded.

See how it signs in

#02 - Signing in

Get your agent through the login, and keep it there.

Passwords live encrypted on your disk and get typed into the form. The agent never sees the value.

login.live.com
Microsoft

Sign in

Use your Microsoft account.

Email or phone number
Next

Every profile has its own address, with no email account behind it. Codes arrive where the agent works.

The six digits are computed inside the profile, from a secret that never leaves it. SMS arrives the same way.

Keys survive a crash and the window closing. Permission bubbles and device pickers get answered instead of stopping the run.

See what the agent gets

#03 - The agent layer

Hand your agent a page it can operate.

The agent looks at the page once, then only at what moved: 13,507 tokens where Playwright MCP needs 2.1 million.

shopify.com/login

First look

31 lines · 935 B

title: Log in — Shopify
[e_122] heading "Log in"
[e_123] heading "Continue to Shopify"
[e_49] button "Sign in with passkey"
[e_25] link "Continue with Google"
[e_28] link "Continue with Apple"
[e_31] link "Continue with Facebook"
[e_34] link "Continue with WhatsApp"
[e_16] button "Continue with email" (disabled)
[e_3] textbox "Email"
[e_39] link "Get started"
[e_208] link "Need Help?"
and 19 more

After one keystroke

4 lines · 182 B

+ [e_277] StaticText "Continue with email"
+ [e_159] image
~ [e_16] button "Continue with email" → no longer disabled
~ [e_3] textbox "Email" → = "ops@northgate.example"

Every step after the first carries the change, not the page.

Press and hold, image grids, sliders, rotation, the spoken one. They work because the input is human.

Every tool call becomes one step with its result and a screenshot. A click that did nothing is its own line.

Read-only mode, address rules, no file system, no payments. Keys and passwords never reach the model.

See what comes out of a run

Videos

Your agent shoots the demo while it works.

The camera moves to whatever just happened, the cursor glides between the real click points, and a voice track with subtitles is generated with it. Five export targets, from the file for your landing page to an upright cut for social. Recording changes nothing about the run: a keystroke lands 0.03 milliseconds later than without it. The same recording also renders as proof, without zoom and with a burnt-in timestamp.

Videos · midjourney-night-run

Video

Four upscales, picked and organised

1:12
  • 0:00Open the wall and wait for the grid
  • 0:14Type the prompt, four images come back
  • 0:31Upscale the two that work, file them for the client

Files this render wrote

  • Video

    MP4 · 1920×1080 · 24 MB

  • Social export

    MP4 · 1080×1920 · 30 fps, the same film cut upright

  • GIF, subtitles, poster

    A preview loop, SRT and WebVTT, an OpenGraph still

Agent integrations

Keep the agent you
already built.

I use it to run about thirty agents that each need their own logged in account. Since every agent needs a separate identity I wanted something where the profile stays the same between runs.

It keeps the session, the cookies and the passkeys per profile. I haven't had to restart the nightly run since June.

Ines Halvorsen

Ines Halvorsen

Automation lead

Kettle

Not taking our word for it?

Ask a model that has never met us, then read the free check on your own browser.

Ask ChatGPT

Ask Gemini

Ask Claude

Ask Perplexity

Run the free check