Last week, I asked my AI agent to scrape 300 job listings from a recruiting site. It broke at row 50.
I retried. Broke again. Different error, different row. Two hours gone - not building something, just fighting a tool that was supposed to save me time.
Sound familiar?
The Real Problem With AI Agents and the Web
Here's the thing nobody tells you when you start using AI agents for real work:
They're incredible at generating code. They're terrible at the web.
Every time you ask an agent to scrape, monitor, or interact with a website, it starts from scratch. Different path. Different failures. Unreliable at scale. And the moment the website redesigns? The whole thing breaks.
web_fetch is basically curl with a nicer name. It can't handle login walls, JavaScript-heavy SPAs, captchas, or dynamic React dropdowns. I've spent more time debugging broken scrapers than actually building things.
Hand-written scrapers are fragile. SaaS scraping tools cost $99/month per website. Manual copy-paste is a waste of human potential.
Then I found Skill Forge - and it changed how I think about AI agents entirely.
The Solution: Explore Once, Reuse Forever
Skill Forge is a meta-skill inside BrowserAct - a free, open-source CLI tool built for AI agents. It's not another scraping SaaS. It's a local tool that turns any website's data or operations into a reusable Skill package (a SKILL.md + Python scripts).
Explore once. Reuse forever.
Instead of your agent re-exploring the same site every run (different path, different failures), Skill Forge does a one-time deep exploration, generates a verified script, and from that point on you just call the Skill. Same stable path. Every time.
Install it with one command:
npx skills add browser-act/skills --skill browser-act-skill-forge
That's it. Free. No signup. No subscription.
🏆 Product Hunt: BrowserAct recently hit #1 Product of the Day and #3 Product of the Week on Product Hunt - proof that developers are hungry for tools that actually let agents work on the real web.
Live Demo: Forging a GitHub Repository Skill
Let me show you exactly how this works with a real example I ran myself: extracting GitHub repository data across multiple repos.
Step 1 - Describe What You Want
I opened VS Code with my AI agent and gave it a plain English description:
I want to extract GitHub repository data. Build a reusable Skill that extracts repository name, GitHub URL, star count, fork count, and last updated date from any repository.
No code. No selectors. No API keys. Just a description.
Step 2 - Automatic Exploration
Skill Forge starts inspecting the site immediately. The first thing it does? Hunts for internal API endpoints. If it finds one, that's 10x more stable than scraping the DOM - and GitHub has them.
- Detected GitHub's internal API structure
- Mapped the request/response patterns
- Identified exact parameters for repository metadata
No intervention from me. It figured this out on its own.
Step 3 - Skill Generation
Once Skill Forge understood the data flow, it generated a complete Skill package:
- SKILL.md - documentation and usage instructions for the Skill
- extract-repo-metadata.py Python script with business parameters extracted as CLI arguments
The Skill is now production-ready. One command to run it. Forever.
Step 4 - Auto-Test (No Human Needed)
A sub-agent automatically ran end-to-end tests on the Skill verified the extraction worked correctly, and if it failed, fixed itself with no intervention needed. It's not just generating code it's validating that the code actually works before handing it to you.
Step 5 - Run at Scale
I ran the Skill on three repositories. Here's the agent executing extraction across all three:
And here's the actual final output:
| Repository Name | GitHub URL | Stars | Forks | Last Updated |
|---|---|---|---|---|
| browser-act/skills | https://github.com/browser-act/skills | 8 | 160 | Jun 26, 2026 |
| microsoft/playwright | https://github.com/microsoft/playwright | 65k+ | 3.5k+ | Jun 28, 2026 |
| modelcontextprotocol/servers | https://github.com/modelcontextprotocol/servers | 15k+ | 1.2k+ | Jun 26, 2026 |
Three repos. Clean structured output. One Skill I can call again tomorrow with different repo names zero re-exploration.
What Makes This Actually Work
Anti-Detection Stealth & Captcha Solving
BrowserAct runs stealth browsers, masks fingerprints, and handles captchas automatically - tested against reCAPTCHA v3, Cloudflare Turnstile, DataDome. Passes checks that standard Playwright fails on.
Persistent Sessions & Multi-Account Isolation
Login once, reuse sessions. Run multiple accounts in parallel without cross-contamination. No repeated 2FA prompts.
Human-in-the-Loop
When a captcha or 2FA genuinely can't be automated, BrowserAct pauses and hands the live browser session to you. You complete the step, the agent resumes.
Proxy Control
Static proxies for stable identity across long-lived sessions. Dynamic proxies for rotating IP on batch collection runs.
Skill Forge vs. The Alternatives
| Approach | Problem | With Skill Forge |
|---|---|---|
| Hand-written scrapers | Break when site changes layout | Forge once, re-Forge to update |
| Agent re-explores every run | Different path, different failures every time | Stable verified Skill, predictable every run |
| SaaS scraping tools | One website, monthly bill ($99–$300+) | Free, runs locally, works on any site |
| Direct bulk execution | Anti-bot triggers mid-run, session drops | Verify the Skill first, then scale safely |
The key insight: don't make your agent explore every time. Forge once then run 500 records or 5,000 with the same stable path.
The Pricing (This Is the Surprising Part)
I expected this to be a SaaS with a monthly bill. It's not.
BrowserAct Skill Forge is 100% free to install and use locally. No subscription. No card. No signup paywall. Proxies are optional add-ons - not required to get started.
Try It
npx skills add browser-act/skills --skill browser-act-skill-forge
⭐ Star the repo on GitHub - readers who star get 500 free credits to try out the dynamic proxy feature.
👉 Get started with BrowserAct →
The web is where real work happens. Login walls, dynamic pages, captchas, anti-bot systems - AI agents hit all of these and fail.
Skill Forge is the layer that fixes it: one exploration, permanent capability.
What website would you Forge first? Drop it in the comments below. 👇
Top comments (0)