Hi everyone! 👋 I'm Tim, founder of Expand, and today we're launching Expand Fetch.

TL;DR: Send a URL, get back Markdown written by Iris, our own vision model. It keeps the tables, icons and hidden JSON other tools drop, and every line links back to where it came from. On our benchmark, agents answer 74% of questions from our Markdown. The next best API gets 55%. Live today, $1 per 1,000 fetches.

https://youtu.be/qQCqpShmrLA

🚨 The problem

Your agent has an IQ of 500, but fails at basic questions about the web.

Web Markdown APIs drop icons, break tables and miss the JSON in the page. So your agent answers wrong, with confidence. And you can't tell whether the page didn't load or the conversion dropped it. The workaround is a full browser per page, which is slow and expensive.

We tried fixing it with heuristics. It's whack-a-mole. The rule that drops a footer deletes a company's address on one page and ships 100,000+ tokens of trash on the next.

🚀 Expand Fetch

One API call gets you:

• 👁️ Iris: our vision model sees the rendered page and writes frontier quality Markdown. Tables stay tables. Navigation and footers go to an appendix.
• 🩻 X-Ray: finds the useful JSON hidden in the page, like coordinates or flight gates.
• 🕰️ Wayback Machine: every fetch is a snapshot you own. Every line links back to its spot on the page.
• 🔦 Highlights: ask a question, get only the passages that answer it. Over 95% fewer tokens.

Plus Crawl for whole sites and WebSync to follow every change.

Or paste this into your coding agent and it sets itself up:

Set up https://expand.ai/start.md

📊 Benchmarks

Existing Markdown benchmarks assume a page has one main content, like reader mode. The web doesn't work like that. Iris sorts every element into preserve, prune or appendix, and on our validation set it beats Jev on F1, 92.7% to 69.6%.

Then the test that matters. We built MDBench with our customers: 1,094 questions on 326 hard pages. Expand is also the only API we tested that returns the hidden JSON.

👶 About us

Expand is an applied research lab. We want to make the web queryable like a database, and found we first need to see it. Iris took us a year of sweat and tears. I've spent 10 years in developer tools, as founding engineer at Prisma and founder of Stellate.

📢 Ask

If you, or anyone you know, has struggled with the quality of web data, including for your own agents, please reach out at tim@expand.ai. Send me a page your agent gets wrong, and I'll show you what Expand returns.

Try it at expand.ai with $10 in free credits, or join the waitlist if you want to query the web like a database.