Loading video...
Video Failed to Load
How to automate web tasks (even when the site doesn't offer an API). Use this with Claude Code, Cursor, OpenCode, OpenClaw, or whatever agent you are using.
27,347 views • 5 months ago •via X (Twitter)
25 Comments

I'm using in the video. It's incredibly easy to create an agent that operates a site. You can define the tasks you want to complete, and then access them through their API. A huge benefit of this: You'll communicate with the site back and forth using structured data. The whole process of running a computer, navigating the site, and figuring out what to do is all hidden behind the scenes. By the way, this works even with sites protected with MFA. Thanks to the @deck_agent team for giving me access to this and partnering with me on this post.

The scary part isn't AI. It's that developers are still ignoring stuff like this. 2 years from now a lot of people are gonna wish they paid attention.

This is cool- Any other examples where you’ve used it? Thinking of trying it for some fintech/investment use cases

Did this on a scraper I’ve been fighting for weeks. Took 4 prompts instead of 4 hours. Chef’s kiss.

The flaky part is selectors. Dynamic sites break everything fast.

Impressive Helpful

feels like headless browser hacks got a new wrapper.

IT developers , coders would be over the moon if they had the same AI assistants in 2015 that are available nowadays

Nice!

@svpino honestly, this is gonna save me so much time. tried doing it manually before and it was a nightmare lol. can't believe i didn't think of automating it sooner!

cursor is a great choice, what kind of web tasks are you automating with it that don't have an api available

What tasks are you automating most? I keep hitting rate limits on scraping financial data for my app.

browser automation gets way cleaner when you add explicit waits for dynamic content. saves you from those random failures when elements load async or js takes forever to render

curious how deck handles session state and auth re-challenges mid-task. that's where our browser-agent flows fall over, not the dom parsing. the api facade is the easy part, it's the stateful login dance that breaks.

yeah the demo part is easy. the ugly part starts when login expires, rate limits kick in, or the same click lands twice and now youve bought two labels. no-api automation gets interesting once side effects are real

The web automation angle without an api is one I keep coming back to. The techniques that work reliably with Claude Code have improved noticeably, though the edge cases still require some manual oversight

how are you handling sites that silently a/b their dom on you? i hit this shipping an agent against a vendor dashboard, selectors broke every couple weeks until we added semantic fingerprinting on top. curious if deck wraps that or just retries.

How do they keep those secrets of yours? Would be very hesitant to upload my passwords to anything that matters to some suddenly popped up startup. Background in fintech gives some credibility. How to rotate keys or passwords to service that does not support automating it?

The underrated part: sites that block scrapers often allow browser automation just fine because it looks like a real user. The agent sees what the user sees -- no API docs needed, no reverse engineering. Works especially well for workflows that mix logged-in state with data extraction. The underrated part: sites that block scrapers often allow browser automation just fine because it looks like a real user. The agent sees what the user sees -- no API docs needed, no reverse engineering. Works especially well for workflows that mix logged-in state with data extraction.

Sounds interesting!

Wow that's expensive

100%. Solve bleeding-neck problems, not vanity problems. Users don’t pay for “cool” - they pay for relief, speed, money, or status. The founder who ships weekly and learns directly from users compounds faster than teams hiding behind planning decks.

no-API automation is where the interesting edge cases live. authentication expiry, DOM mutations, rate limits that look like captchas. the agent has to recover without asking for help.

What about captcha?

the secret is finding the right selectors. once you've got those, the agent does everything else without you watching
