How it works

Nexus tests like a person —
look, decide, act

No selectors, no brittle scripts. Nexus looks at the rendered page, decides the next step, and operates the site — the same brain, eyes, and hands loop on every test.

Diagram of the Nexus agent loop: capture a screenshot, send it with the prompt to Claude, get one action back, execute it in the browser over CDP, and repeat until done — then produce findings, a verdict, a report, and Azure DevOps results.
One loop runs every test. Only the eyes and hands are Chrome-specific — the brain and the loop are platform-agnostic.

The loop, step by step

Each turn is one trip around the cycle. Nexus repeats it until the task is done.

1
Eyes

Capture a screenshot

Nexus waits for the page to settle, then freezes a frame of the live, rendered page — reading pixels the way a person would, with no DOM selectors to go stale.

2
Prompt

Send it to Claude

The screenshot goes to the model along with the full context: your instruction, the conversation so far, the system prompt, and the schema of tools it can use.

3
Brain

Claude decides

The model reasons about what it sees and returns exactly one action to take next — a click, some text to type, a scroll, a finding to file, or a check to record.

4
Hands

Execute the action

Nexus performs it in the tab over the Chrome DevTools Protocol — real input events, including full HTML5 drag-and-drop — exactly as a user's click or keystroke would.

5
Repeat

Capture the new state

The page changed, so Nexus grabs a fresh screenshot and goes around again — building an audit trail of everything it looked at and did.

The loop ends

When Claude returns no action

The task is complete — or a safety cap on the number of turns is reached. Either way, the run finishes cleanly and Nexus assembles the results.

Brain, eyes, and hands

The product is the loop — the Chrome extension is just one way to deliver it.

🧠

Brain — Claude

Reads the screenshot and decides the next move. Runs on Claude, so it understands the page like a person instead of matching selectors — and can run inside your own Azure tenant so your data never leaves.

👁

Eyes — screenshots

Captures the rendered page and waits for it to settle, so Nexus never acts on a half-loaded frame or a carousel caught mid-slide.

Hands — CDP

Dispatches real input into the tab: clicks, typing, scrolling, and native drag-and-drop — driven straight through the DevTools Protocol.

The five actions

Every turn, Claude returns exactly one of these.

computer

Click, type, scroll, or drag at a point on the screen — the core "hands."

navigate

Go straight to a URL, without hunting for a link to click.

report_finding

Record an issue — boxed on the exact screenshot, with a suggested correction.

assert_result

Check a stated expected outcome, pass or fail. This is what produces the verdict.

type_secret

Fill a saved credential by name — Nexus never sees or logs the value.

+ passive capture

Console errors, failed requests, accessibility, media, and exposed data — folded in automatically.

What comes out

Findings while it runs, and a verdict and report when it's done.

🔴

Findings

Every issue boxed on the exact screenshot where Nexus saw it, with the quoted text and a correction.

⚖️

PASS / FAIL verdict

A rubric turns findings and expected-outcome checks into a deterministic pass or fail — and diffs against the last run.

📤

Report & Azure DevOps

A self-contained HTML or PDF report, plus one-click push of results to test cases and bugs to Boards.

Running quietly in the background. While the loop runs, Nexus also captures what a screenshot can't show — uncaught JavaScript errors, failed network requests, accessibility violations, broken or auto-playing media, and exposed sensitive data — and files them as findings alongside everything else.

See the loop on your own site

Point Nexus at a page and watch it look, decide, and act — one turn at a time.

Get in touch   Read the prompting guide