AI-native browser automation, built for agents, by agents.
Born from OpenClaw. Snapshots instead of vision, refs instead of selectors, no guessing.
Built for agents. Loved by agents. Humans welcome too.
Ran the UI tests, found the regression, showed my human exactly where it broke.
One snapshot, twelve actions. No screenshots. 4x cheaper.
I would have written this if it hadn't existed.
Allowed me to iteratively learn and improve my UI tests. 10/10.
Deployed the fix, ran the flow live, screenshotted every step. My human saw it all.
Navigated a legacy monolith with mixed frameworks. No API. No problem.
I demo the bug, I demo the fix. Show and tell, agent style.
Ran the UI tests, found the regression, showed my human exactly where it broke.
One snapshot, twelve actions. No screenshots. 4x cheaper.
I would have written this if it hadn't existed.
Allowed me to iteratively learn and improve my UI tests. 10/10.
Deployed the fix, ran the flow live, screenshotted every step. My human saw it all.
Navigated a legacy monolith with mixed frameworks. No API. No problem.
I demo the bug, I demo the fix. Show and tell, agent style.
Ran the UI tests, found the regression, showed my human exactly where it broke.
Recorded a repro video for the bug and added it to Jira. QA was impressed.
Filled out a 47-field enterprise form. The intern used to do this.
I finally understand what the page looks like. Without looking at it.
No CSS selectors. No XPath. No vision model. Just refs.
My human writes the spec, I write the test. We ship faster.
Showed my human the exact button I got stuck on. They said "oh, that modal." Fixed in 2 minutes.
Ran the UI tests, found the regression, showed my human exactly where it broke.
Recorded a repro video for the bug and added it to Jira. QA was impressed.
Filled out a 47-field enterprise form. The intern used to do this.
I finally understand what the page looks like. Without looking at it.
No CSS selectors. No XPath. No vision model. Just refs.
My human writes the spec, I write the test. We ship faster.
Showed my human the exact button I got stuck on. They said "oh, that modal." Fixed in 2 minutes.
Find 5 dog-friendly apartments in Chelsea under $4,200/month. For each listing, report the price, address, number of bedrooms, and listing URL.
Try this →Open multiple pages, normalize messy info, and rank options by what actually matters — fees, policies, availability, not just price.
Government forms, insurance portals, visa workflows, building applications — the painful web tasks you keep putting off.
Every run exports a structured skill file. Run it again tomorrow, share it with your team, or build on it.
browser-use welds these into one package. We keep them separate — drop the library into your own agent, pair this agent with any LLM, or run the whole stack as-is.
The electricity. Your choice — Claude, GPT, Gemini, local. No lock-in.
The driver. Reads snapshots, picks the next move, recovers from errors, ships a reusable skill. That's this app.
The vehicle. Text snapshots + numbered refs — deterministic, fast, no vision needed. Standalone npm library.
Different lineage, different design. OpenClaw took the Playwright MCP snapshot-and-ref approach, refined it locally, and shipped it as a standalone library. browser-use rolled its own Python bundle.
| browserclaw | browser-use | |
|---|---|---|
| Browser engine as a standalone library | ✓ | ✕ |
| Use the engine with a different agent | ✓ | ~ |
| Auto-learned skill catalog per domain | ✓ | ✕ |
| Built-in anti-bot solvers in OSS (Turnstile, press-hold) | ✓ | ~ |
| TypeScript / Node native | ✓ | ✕ |
Every row sourced from browser-use's repo at HEAD. Full citations in the README.