← Skills

Give your coding agent a real browser

2026-07-01

Coding agents are good at reading and writing code. They're much worse at telling you whether the thing they just built actually works in a browser — because by default, they can't open one. "Does this page load?" and "did that form submit?" are questions that need a browser, not a source-code read.

This skill gives your coding agent that browser.

What this skill does

web-browse lets your agent drive a real browser session: navigate to a URL, click and fill in elements, wait for content to appear, extract text or structured data, and take screenshots it can read back and reason about. It runs in two modes:

  • Cloud mode — an isolated browser session spun up on demand, no local setup, good for scraping, research, or checking a public page.
  • Local mode — your agent drives your actual browser tab, so it can work with sites that need your real login session, without you handing over a password.

Either way, your agent gets the same primitives it would need to test or research anything on the web: navigate, click, type, wait, extract, screenshot, read console/network logs.

…have your coding agent do it, on ThinkRun

This skill is powered by ThinkRun, which is what makes "give the agent a browser" safe and useful rather than a security problem waiting to happen:

  • Your login stays yours. Local mode runs in your own browser session — nothing hands your credentials to a third-party service, and your agent is built to never type a real password; sign-in stays something you do yourself.
  • Every session is a recording. Whatever your agent does — cloud or local — ThinkRun captures it as a replayable trace, so "the agent visited some pages" isn't a black box. You can review exactly what it clicked and when.

The two modes cover the two situations that come up constantly: a public page you just want checked or scraped (cloud, zero setup), and an internal tool or app behind a login that only your browser can reach (local, using your real session).

One skill, a bigger idea

Giving an agent a real, recorded browser is the foundation the rest of the ThinkRun skill library sits on — from structured UX audits to filling out dashboards that have no API. This is the general-purpose version: point it at any page and let your agent take it from there.

Install

Download the file and drop it into your agent's skills directory (for Claude Code: .claude/skills/web-browse/):

So the layout is .claude/skills/web-browse/SKILL.md. Then just ask your agent to "check if this page loads", "scrape this page for me", or "fill out this form."

Prerequisite: this skill drives ThinkRun, so your agent needs it installed first — the Chrome extension for local mode (real, logged-in browser), or just npx @thinkrun/mcp / the CLI for cloud mode. Full setup: docs.