Our Approach

We operate below
the interface.

Instead of hooking into the browser's internals, we interact with software the same way a human does: through the operating system. Real pixels. Real input devices. Real timing. Real residential IPs matched to each session's geo.

01

Real Chrome, not patched Chromium

We run unmodified Chrome with full rendering pipelines, GPU compositing, and all standard browser behaviors intact. No automation hooks. No DevTools injection. No detectable substrate.

02

OS-level input, not dispatched events

Inputs flow through the operating system’s native input stack — the same path a physical keyboard and mouse use. The browser cannot distinguish our input from human input because the delivery mechanism is identical.

03

GPU-backed rendering, not headless mode

Full pixel rendering through the graphics pipeline. WebGL, Canvas, and every visual fingerprinting surface behave exactly as they would on a real machine because they are running on a real graphics stack.

04

Isolation without virtual machines

We achieve per-session isolation without the overhead of full VMs. This keeps cost per session low enough to be viable at scale — the economics that killed every previous attempt at this approach.

05

Recovery-first, not script-first

Real-world software is nondeterministic. Modals appear. Pages stall. Layouts shift. We build for recovery and adaptation, not brittle step-by-step scripts that break on the first unexpected state.

06

Clean network identity, not rotating proxies

Every session gets a residential IP matched to its fingerprint’s geo. No datacenter ranges that get blocklisted on sight. No geo mismatches that trip Cloudflare in milliseconds. The network layer is as real as the browser layer — because a perfect browser on a flagged IP is still a blocked session.

How It Works

The execution pipeline

Every task flows through five stages: from the AI agent's instruction to a completed action inside human-only software. No DOM hooks. No synthetic events. Just real execution.

01
Receive
Task Received
AI agent sends structured instruction
02
Spawn
Environment Ready
Isolated session with real Chrome, GPU & residential IP
03
Execute
Action Performed
OS-level input drives the interaction
04
Observe
State Verified
Visual + DOM state reconciliation
05
Return
Result Delivered
Data returned or recovery triggered
DOM Automation — CDP hooks, synthetic events, breaks on legacy
Execution Runtime — OS input, real rendering, task complete

The Architecture Problem

Existing tools assume the browser cooperates.

Every browser automation tool today (Playwright, Puppeteer, Selenium, RPA frameworks) operates above the interface. They hook into DevTools, inject scripts, and dispatch synthetic events. This works when software cooperates. Most real-world software does not.

DOM Automation
  • Controls browser via DevTools Protocol
  • Uses patched Chromium with automation hooks
  • Dispatches synthetic DOM events
  • Datacenter IPs that get blocklisted on sight
  • Fingerprint-to-geo mismatches trip CDN detection
  • Assumes stable, cooperative interfaces
  • Detectable by fingerprinting and behavioral analysis
Execution Runtime (Our Approach)
  • Operates below the interface at the OS level
  • Uses real Chrome with real rendering paths
  • Drives native input devices through the kernel
  • Residential IPs matched to session fingerprint geo
  • Consistent identity per session — no random rotation
  • Embraces nondeterminism with recovery-first design
  • Interaction patterns indistinguishable from human usage

The Distinction

Interfaces vs. Interactions

Browser automation tools automate interfaces: DOM nodes, selectors, events. But real work happens at the interaction level. Focus, timing, context, recovery. That's where we operate.

Interface Automation
Selectors
querySelector, XPath, CSS targets
DOM Nodes
Direct tree manipulation
Synthetic Events
Programmatic event dispatch
Scripts
Injected JavaScript execution
Determinism
Expects predictable outcomes
Interaction Execution
Focus & Context
Real window focus, z-order awareness
Timing & Cadence
Human-natural input timing
State Awareness
Visual + DOM state reconciliation
Error Recovery
Retry, adapt, recover from failures
Side Effects
Real-world consequences, not assertions
“Playwright automates interfaces.
We automate interactions.”

Where We Fit

The execution stack

Layer 3
AI / Business Logic
LLM planners, workflow engines, agents
Layer 2
Our Execution Runtime
Real browsers, real inputs, clean IPs, recovery
Layer 1
Human-Designed Software
No APIs, hostile, legacy, regulated

DOM automation tools live above the interface. We live below it. That's why they can't evolve into us. Playwright stays for tests and friendly software. We exist for everything else.

Why is this the right time to build this?

Why Now →