Skip to main content
Back

Hyperbrowser Ships a Claude Code Skill That Maps Websites Before AI Agents Browse

Hyperbrowser's /web skill generates a structured web.md map of any site, giving AI coding agents the context they need to navigate it reliably

Repo, Full Access
  • /web skill: Hyperbrowser released a Claude Code slash command that crawls any website and generates a structured web.md map for AI agents.
  • Problem solved: AI agents navigating websites blind waste tokens and fail tasks; web.md gives them full site structure before they start.
  • How to install: Add Hyperbrowser MCP via .mcp.json, drop SKILL.md into .claude/skills/web/, restart Claude Code -- available now.
  • Pricing: Free tier available; credits cost $0.001 each; a 100-page crawl costs well under $1 at $0.10/hr browser time + $0.001/page scrape.
  • MCP toolkit: The underlying Hyperbrowser MCP server (768 GitHub stars) exposes scraping, crawling, structured extraction, and AI browser agents.
  • Traction: The launch post hit 74K views and 649 bookmarks, signaling strong demand for more reliable web agent tooling.

AI agents that browse the web have a dirty secret: they're flying blind. They land on a page with no idea what's on the next one, what flows exist, or which buttons actually do something important. The result is a lot of wasted tokens, wrong clicks, and failed tasks. Hyperbrowser just shipped a practical fix for this: a Claude Code skill called /web that crawls any website and produces a web.md file -- a structured, LLM-friendly map of every page, action, flow, and interactive element on the site.

The idea is simple but the impact is real. Instead of sending your agent into a website cold, you run /web first. It hands the agent a complete map before it takes a single step.

The blind navigation problem

When you build an agent that needs to interact with a web app -- say, to fill out a form, trigger a workflow, or extract data from a dashboard -- the agent has to figure out the site's structure on the fly. It scrapes one page, guesses what to click next, and often ends up in a dead end or loops back to somewhere it's already been. This isn't a model intelligence problem; it's a context problem. The agent simply doesn't know the map.

Hyperbrowser specifically targets AI agent applications, optimizing infrastructure for autonomous browser agents with features like CAPTCHA bypass, session management, and AI agent optimization. The /web skill is the latest layer on top of that infrastructure -- a pre-flight reconnaissance tool for agents.

How it works

The /web skill is a Claude Code skill -- a markdown file dropped into .claude/skills/ that teaches Claude Code a new slash command. Skills are a relatively new pattern in the Claude Code ecosystem: they're essentially system-prompt fragments that activate when you invoke a specific command, giving the agent a precise set of instructions for a well-defined task.

When you run /web and point it at a URL, it uses the Hyperbrowser MCP Server, which provides tools to scrape, extract structured data, and crawl webpages, as well as easy access to general-purpose browser agents like OpenAI's CUA, Anthropic's Claude Computer Use, and Browser Use. The crawl produces a web.md file that documents the site's structure in a format optimized for LLMs to consume.

The setup involves three steps:

  1. Add the Hyperbrowser MCP to your project by creating a .mcp.json at the root with your API key
  2. Create the skill directory at .claude/skills/web/ and drop in the SKILL.md file from the Hyperbrowser examples repo
  3. Restart Claude Code and invoke /web <url>

The MCP config looks like this:

// .mcp.json
{
  "mcpServers": {
    "hyperbrowser": {
      "command": "npx",
      "args": ["-y", "hyperbrowser-mcp"],
      "env": {
        "HYPERBROWSER_API_KEY": "YOUR_HYPERBROWSER_API_KEY"
      }
    }
  }
}

What the MCP server actually gives you

The Hyperbrowser MCP server (768 stars on GitHub) exposes a full toolkit that the /web skill draws on under the hood:

  • scrape_webpage -- extract formatted content (markdown, screenshots) from any page
  • crawl_webpages -- navigate through multiple linked pages and extract LLM-friendly content
  • extract_structured_data -- convert messy HTML into structured JSON
  • browser_use_agent -- fast, lightweight browser automation
  • claude_computer_use_agent -- complex tasks using Claude's computer use capability
  • openai_computer_use_agent -- general-purpose automation using OpenAI's CUA model

Browser automation agents include specialized cloud browser agents from OpenAI's CUA, Anthropic's Claude, and Puppeteer, to perform precise, reliable, and complex web interactions based on detailed instructions.

Where this shines -- and where it doesn't

The /web skill is most valuable when you're building an agent that will interact with the same site repeatedly. Generate the map once, commit it to your repo, and every subsequent agent run starts with full context. Think:

  • Automating SaaS workflows (form submissions, dashboard interactions, report exports)
  • Writing end-to-end tests for your own web app before you know all the routes
  • Onboarding a coding agent to a third-party tool it's never seen
  • Scraping multi-page sites where link structure isn't obvious from the homepage

It's less useful for highly dynamic sites where the structure changes frequently, or single-page apps where most content is loaded client-side behind authentication walls. The map is only as good as what the crawler can reach without logging in.

Pricing

Hyperbrowser offers a free tier with starter credits to experiment without spending, a Startup plan at $30/month, and a Scale plan at $100/month with more credits and pay-as-you-go overages. Credits are priced at 1 credit = $0.001, or 1,000 credits = $1.00. Crawling a site for web.md generation consumes browser session time at $0.10/hour plus scraping costs at $0.001/page -- meaning a 100-page crawl costs well under a dollar in most cases.

The bigger pattern

The /web skill is part of a broader shift in how the industry thinks about agent reliability. The current default -- throw an agent at a task and hope it figures out the environment -- produces brittle, expensive pipelines. The emerging alternative is to invest in upfront context: give the agent a map, a schema, or a plan before it starts acting. web.md is that map for the web.

Hyperbrowser started as a response to a core pain many automation engineers know well: running headless browsers at scale without managing the infrastructure yourself. The founders built it after grappling with scaling, proxies, and bot defenses on their own AI-heavy projects, then packaged those solutions into a cloud service. The /web skill is the most developer-facing expression of that yet -- less infrastructure, more workflow.

With nearly 74,000 views on the launch post and 649 bookmarks, the signal is clear: people building web agents are hungry for anything that makes them less brittle. A pre-generated site map is a low-effort, high-leverage intervention, and the fact that it plugs directly into Claude Code's existing skill system means there's almost no adoption friction.