Skyvern (YC S23)

A company I talked to this week has 1,000+ native app integrations. CRMs, ERPs, accounting systems, all of it. They still needed browser automation because one supplier portal has no API.

Your automation is getting flagged before it sends a single request. IP rotation doesn't work anymore. Bot detection lives in the browser fingerprint now: hardware values, TLS patterns, canvas rendering. A unique IP with a cloned fingerprint is just 20 copies of the same browser

Talked to 47 customers this quarter. Most common complaint about their previous vendor? Not the price. Not missing features. "It breaks." Reliability is the most underrated competitive moat in software.

The three most interesting builders I talked to this week: A neurology practice co-founder. An appraisal firm owner. An ice machine leasing operator. None of them engineers. All running production automation workflows. The definition of "builder" is changing.

We changed the default AI model for a customer. A story about the chaos that followed

There is no better feeling than optimizing a few tables and seeing your CPU usage drop

New landing page alert!!! What do you guys think??

Best sales framework I've seen: sell the cost of the problem before you ever show a feature. Buyers don't care what your product does. They care what their life looks like after the problem is gone.

🚀🚀🚀 We're excited to launch Skyvern MCP — give your AI assistant superpowers to browse the web, fill out forms, extract data, and run multi-step automations. Works with OpenClaw, Claude Code, Codex, Cursor, or your custom agent. 🎯 Get started at app.skyvern.com or check out the open source repo at github.com/Skyvern-AI/sky…. Skyvern MCP connects your favorite AI assistant to a real cloud browser through the Model Context Protocol. Instead of writing Selenium scripts or wrestling with CSS selectors, you just tell your AI what to do in plain English — and Skyvern handles the rest. ⚡ Setup takes 30 seconds. Seriously. No Python. No pip install. No local server. One line of config, and your AI assistant can browse the web: claude mcp add-json skyvern '{"type":"http","url":"api.skyvern.com/mcp/","headers":{"x-api-key":"YOUR_API_KEY"}}' That's it. Your AI now has access to 33 browser automation tools across 6 categories. ❓ How is this different from other browser automation tools? Traditional browser automation (Selenium, Playwright, Puppeteer) requires you to write code, manage selectors, and handle every edge case manually. Skyvern MCP flips this on its head: 🗣️ Natural language, not code — Say "Submit this" instead of writing document.querySelector('#btn-submit-form-v2').click() 🧠 AI-powered extraction — Ask "extract all job listings with title, company, and salary" and get back clean JSON ✅ AI validation — Check conditions like "is the user logged in?" and get a true/false answer 🔄 Reusable workflows — Chain actions into multi-step automations you can run again and again ☁️ Cloud browsers — No local browser needed. Skyvern runs browsers in the cloud with geographic proxy support 🤔 What can you actually do with it? Here are real use cases we see every day: 📊 "Go to Hacker News and get the top 10 posts with titles and scores" — Your AI opens a browser, navigates, extracts structured data, and returns it to you 📝 "Fill out this government form with my business details" — Multi-page form automation with intelligent field detection 🧾 "Log into my vendor portal and download last month's invoices" — Secure credential-based login + file download 🔍 "Search the Secretary of State website and verify this business registration" — Multi-step research across dynamic pages 💼 "Find remote Python jobs on Indeed paying over $150k" — Navigate, filter, extract, all in one conversation 🛠️ 33 Tools. 6 Categories. Infinite Possibilities. 🖥️ Works with all the tools you already use OpenClaw Claude Desktop Claude Code Cursor Codex Any MCP-compatible client

You know your prompt is good when... your agents run for >1 hr Here are 5 tips to make your agents run longer (#5 is really funny) 1. Give it a clear and concise goal (or multiple) goals to accomplish Good examples: "Implement a feature in Skyvern that lets people select different proxy locations after launching a browser session" Bad examples: "Implement proxy locations for skyvern" 2. Give it clear context so it can understand WTF you mean Good example: "We have a browser sessions page where users can create new browsers and use them for automations. When a user creates a new one, they should be prompted for a proxy location" Bad example: NO CONTEXT! 3. Give it a way to know it successfully did the job Good example: "After running the browser session, connect to it with Skyvern's MCP and make sure it matches the locations that you previously selected. Run this test with a few locations such as Canada, Japan, and Germany Bad example: NO SUCCESS CRITERIA! Your agent will make mistakes, and will have no ways to fix them! 4. Ask it to add tests for the future Good example: "Add a smoke test to the test suite that makes sure selecting a proxy location in the UI actually works Bad example: NO TESTS! 5. Ask it to do a self-review Good example: "Review the code as if your top competitor, CODEX, wrote the code. Scrutinize accordingly. Make sure they didnt' miss any changes, and didn't make any incorrect changes and output them with a severity score (in a nicely grouped and formatted message): SEV-1: Critical, SEV-2 Bad, SEV-3 OK, SEV-4 Nice to have" Bad example: NOTHING! You will ship slop this way! Want to guarantee this happens every time? Put all this in your claude.md and watch your productivity 10x overnight

Claude just tried to introduce some bad code to our codebase Claude: const client = await getClient(credentialGetter, "sans-api-v1"); Me: - wtf is "sans-api-v1" - are you writing bad code?

Claude code has changed how our company operates Before? Everyone spent a ton of time manually shepherding things along Now? Claude code is at the epicenter of it all Customer reports a bug? -> Claude code fixes it Customer has a question? -> Claude code answers it We have a question about how something works? -> Claude code answers it It's strictly better than any point solution out there, and hints at the rapture to come Want to learn how it works? Put CLAUDE below and if I get enough excitement I'll post a guide on how we used claude code to scale to compete with 20x better funded companies with only 6 employees

Working with interns / co-op students is really inspiring. So much raw potential and optimism is infectious.

Day 5 - New Pricing Model Automation pricing is usually optimized for the vendor, not the user. You pay per step, per action, per minute, or per obscure unit that only makes sense if you already understand the internals of the system.

Day 4 - Upload your SOPs directly into Skyvern Most companies already know how their work gets done.

We shipped a "Show don't tell" feature in Skyvern. Instead of telling a browser automation agent what to do, you can now show it by demonstrating a workflow once in a real browser session. Why?

Skyvern's new SDK lets you interplay Playwright with cool new AI features You don't have to pick one or the other anymore - you can use both with one unified SDK Want to get started? Just run uv add skyvern and you're good to go!

Launch Week day 1 - Automation copilot Automation Copilot reduces time-to-automation. Not by making automation “easier to start,” but by making it faster to: 1. build 2. modify 3. and maintain real, production browser workflows.

Are they eventually going to make us speed run super mario 64 to access a website?

Nothing more invigorating than doing a 10pm customer call with a customer you love