Once Claude can reach the live web through an MCP server, the next question is always the same: what do I actually build with it? I made an 88-second video of eight projects. Each one is a single prompt. This post has every prompt, the tool Claude picks for it, and why that tool fits the job better than "just scrape the page".
Setup
The examples use CrawlForge MCP, which I build. In Claude Code it is one command:
claude mcp add crawlforge -s user -e CRAWLFORGE_API_KEY=your_key -- npx -y crawlforge-mcp-server@latest mcp
It also works in Claude Desktop, Cursor and any other MCP client. You don't name tools in the prompts below. Claude reads the tool descriptions and picks one.
A note on honesty: the output panels in the video are example data. The tool names, arguments and credit prices are real.
1. Research assistant
"Compare the top open-source vector databases for 2026. Cite every claim."
Tool: deep_research (10 credits), called with maxUrls: 25 and enableConflictDetection: true.
Asking Claude to search and then scrape ten pages takes ten round trips, and every page dump lands in your context. deep_research expands the query, fetches and de-duplicates the sources, scores each one for credibility and flags places where sources disagree. That last part is the useful bit for a comparison post: if two benchmarks give different numbers for the same database, you see that before you publish.
2. Competitor price monitor
"Watch 5 competitor pricing pages every hour. Tell me when a price changes."
Tool: track_changes (3 credits per page per check), first with create_baseline and then with compare, scoped by selector: .pricing.
The selector is what makes this work. Without it, a rotating testimonial or a cookie banner counts as a "change" and you get alerts all day. With it, you compare only the pricing block and get back a change percentage and a diff. track_changes doesn't schedule itself. Run the compare from cron or a scheduled agent, or use the Monitors page in the dashboard if you don't want to write code.
3. Lead list builder
"Find 20 dentists in Austin and flag the ones with a weak website."
Tools: search_web (5 credits) with limit: 20, then extract_metadata (1 credit each) on the 20 homepages. 21 calls, 25 credits.
"Weak website" sounds subjective, but metadata gets you most of the way: a missing meta description, a default title like "Home", or no viewport tag (so the site isn't mobile-friendly). Those are things an agency can actually pitch on. extract_metadata is the cheapest tool for this because it skips the page body entirely.
4. Chat with any docs
"Crawl my framework's docs into clean text I can embed for a support bot."
Tools: map_site (2 credits) to list every URL from the sitemap and links, then crawl_deep (4 credits) with max_pages: 100, max_depth: 3 and content extraction on. 2 calls, 6 credits.
Mapping first means you see the URL list before you crawl, so you can drop changelogs, blog posts or old-version paths that would pollute the embeddings. The crawl then returns main content with navigation and footers stripped, which is what you want to chunk.
5. Market research from Reddit
"What do people complain about in project management apps? Check Reddit."
Tools: reddit_search (5 credits) in mode: comments, then analyze_content (3 credits) for sentiment and topics. 2 calls, 8 credits.
reddit.com blocks scrapers, so a general scrape tool tends to come back empty or blocked. reddit_search goes through community archives instead. Searching comments rather than posts matters here: complaints mostly live in replies ("we left X because..."), not in thread titles.
6. Product data
"Get the price, stock and every variant from this product page."
Tool: extract_embedded_state (2 credits), with find: "price".
Most modern storefronts ship their data as JSON inside the page: __NEXT_DATA__ on Next.js, the product JSON on Shopify, Nuxt and Redux state elsewhere. Reading that payload gets you exact values, every variant included, without an LLM guessing from rendered HTML. If a field is there, it's correct. If it isn't, you get nothing rather than a confident hallucination.
7. Brand style guide
"Pull the brand colours, fonts and button styles from this website."
Tool: scrape with formats: ["branding"] (2 credits).
Pulling hex codes out of a stylesheet gives you a palette with no meaning attached. The branding format renders the page and reads computed styles, so it can tell you which colour is the primary button background and which is body text, plus fonts, the logo and grouped CSS tokens.
8. Logged-in workflows
"Log in to my supplier portal and pull this month's invoices into a CSV."
Tool: browser_session, priced per operation: open (3), snapshot (1), three actions (fill, fill, click at 1 each), read (2), close (1). 10 credits for that run.
One-shot browser tools close the browser when the call returns, so the login is lost. A session stays open across calls. The snapshot lists the interactive elements with stable refs (@e1, @e2, ...), Claude acts on those refs, and if it picks the wrong button you've lost one 1-credit action, not a whole chain.
The pattern
What the eight projects have in common
extract_metadata for lead scoring, embedded state for product data, map_site before a crawl. A full browser render is a last resort, not the default.deep_research call returns a synthesis instead of ten raw pages in Claude's context window.
Which of these would you build first, or what's missing from the list? I'd like to hear what people are wiring Claude up to.
Try it with 1,000 free credits
This article was written with the help of AI and reviewed by the author.
Top comments (0)