chrome-agent: Verified Browser Actions for Agents

Where this fits

You're building an agent that needs to fill a form, scrape a page, or take a screenshot — but you can't trust the agent's own claim that it worked. The page might have changed, the selector might have missed, or the click landed on an overlay. chrome-agent solves this by giving every browser action a structured verdict: did it actually happen, why, and what's next. It turns browser automation from a blind guess into a verifiable step in your agent's loop.

Why agents benefit

• Verified outcomes: Every mutating action returns verdict, verdict_reason, and next — so your agent knows if a click, fill, or select truly succeeded, not just that a command was sent.
• Delivery interception detection: When a click targets a coordinate, chrome-agent hit-tests first and reports which element actually received the event (intercepted_by), exposing overlay or z-index problems immediately.
• Read-back for writes: After filling, selecting, or checking, it reads the field's actual value through the same window (observed_after_ms) and returns value.actual vs. value.requested — no more silent failures on dynamic forms.
• Page-scoped device emulation: Emulate viewports, network conditions, and device metrics for responsive checks without spinning up separate browser contexts.
• Console and network introspection: Capture console errors, failed requests, and network activity per action, giving agents diagnostic signals to recover from broken flows.

Practical scenarios

Scraping a paginated table with retry logic

An agent needs to extract rows across multiple pages. chrome-agent returns structured page content and a verdict on each navigation, so the agent can detect a stale DOM or a blocked request and retry with a different selector or wait condition — without hallucinating data.

Filling a multi-step checkout form

Each fill and select returns the actual value the page kept. If the form resets a field or a validation error clears it, the agent sees value.actual mismatch and can adjust its next action, rather than blindly submitting an invalid form.

Responsive testing across device profiles

The agent sets a mobile viewport, captures a screenshot, and checks window.innerWidth via console. With delivery interception, it can confirm a tap on a hamburger menu actually toggled the menu, not a hidden overlay — all while logging console errors for regression.

Add it to your agent workflow

Install the skill and add it to your agent's toolset:

npx skills add sderosiaux/chrome-agent --skill chrome-agent

Then use it in your agent code (example with Claude-style tool use):

const result = await agent.tool('chrome-agent', {
  action: 'click',
  selector: '#submit',
  wait_after_ms: 500
});
console.log(result.verdict); // 'success' or 'failure'
console.log(result.delivery); // element that received the event

For more context, see the skill page on OpenAgentSkill.

Compare before adopting

Before you standardize on chrome-agent, compare it against alternatives like Playwright MCP, Puppeteer-based skills, or raw CDP wrappers. Evaluate: does it return structured verdicts on every action? How fresh is the codebase (check last commit and issues)? Does it support your target browser versions? And does its action schema fit your agent's tool-calling format? For simple one-off scraping, a lightweight HTTP client might suffice — but if you need reliability and introspection, chrome-agent's design is worth the extra dependency.

Why it is worth tracking

chrome-agent is actively maintained (v0.14.0, 80 stars, regular commits) and its focus on verified outcomes aligns with the industry shift toward agentic reliability. The pattern of returning verdict + actual values is a strong signal for any agent tooling. If you're building agents that touch the web, track this skill — it's a candidate for your default browser automation layer, and its evolution will likely influence how other tools handle action verification.