Skyvern

What Skyvern is

Skyvern is an AI browser automation platform. Its homepage headline is "AI agents to automate workflows on any website", and the problem it describes is familiar to any operations or finance team: people log into dozens of portals to download invoices, copy and paste data, and fill in forms. Skyvern says it automates these workflows natively in the browser, just like a human would. It is aimed at operations and finance teams that live in vendor portals, and at developers who want to call that capability from code. The site names Y Combinator as a backer, shows a copyright line for Ikonomos Inc., and says it is open source ("Check us out, we're open source"). The pages do not give a headquarters location.

What you build and how

The site says there are many ways to build, and every approach "ends in a production-ready workflow". The homepage lists five.

• Copilot Chat. You describe the job in plain English and the AI Copilot turns the instructions into a structured workflow.
• Upload SOPs. The homepage's phrase is "turn docs into workflows", so a written procedure becomes the starting point.
• Record Browser. "Show once, run forever": you perform the task in a browser and replay it.
• Visual Builder. "Drag, drop, refine", for people who want to see and edit each step on a canvas.
• Native SDKs. A code-first route with Python and TypeScript SDKs. The code sample on the page opens a cloud browser, goes to a page, logs in with a stored credential and runs a natural-language task such as checking out an order.

What makes this different from a classic selector-based RPA script is the engine. The homepage contrasts Skyvern with Playwright and Selenium, saying selector scripts break when the interface changes and need constant engineering upkeep, whereas Skyvern "uses vision instead of brittle selectors" and its workflows survive UI updates. A blog article on the site adds detail: Skyvern adds AI-driven actions to Playwright and uses vision-language models to match a request to elements on the page. A workflow can mix normal Playwright actions, natural-language actions and an AI fallback when a selector stops working. Against other RPA tools, the homepage says Skyvern blends LLM reasoning with executable code to generate deterministic, repeatable workflows.

Hard parts of real portals. The homepage lists CAPTCHA support (solved natively, with no third-party solving services), two-factor authentication including TOTP, and proxy networks with country, state or zip-code level targeting across residential and datacenter networks. It also lists explainable AI: every run has built-in summaries of each action the AI takes and why.

What it produces. Data extraction returns results in any schema you choose, including CSV or JSON. The use-case list on the homepage names six jobs: document processing (download, parse, classify and route documents from any portal, with OCR and structured output), accounts payable (pulling invoices from hundreds of vendor portals behind logins, 2FA and CAPTCHAs), data extraction (pricing, product catalogs, public records), form filling at scale (applications, registrations and compliance forms), lead generation and outreach, and testing and QA, described as browser testing that adapts to UI changes.

Connections. The integration strip shows Zapier, Make, Clay, n8n and Workato, with a link to the full list. For developers the page lists Python and TypeScript SDKs, webhooks and event streaming, readiness for the Model Context Protocol (MCP), and the option to run entirely on-premise with Docker containers.

What the end user sees. For an operations team the result is a workflow that runs without a person logging in: invoices land in a folder, a form gets submitted, a JSON file arrives. A developer sees an API call that returns data.

What the pages say about scale, security and cost

The homepage carries several numbers, all of them company claims. It says 30,000+ users and customers, 500+ enterprise users, 10M+ workflows run, a 99.9% uptime SLA and 23,100+ GitHub stars (the navigation shows 23.1K). A row of logos sits under "Giving time back to teams at": CarEdge, Valence Intelligence, rapidAND, Legion Health, ONLAYER and Pathmonk. The page does not say what each logo's relationship to Skyvern is.

On security, the homepage badges read HIPAA Compliant, SOC2 Compliant and Y Combinator backed, and a later "Enterprise-grade security by default" block shows SOC2 Type II, HIPAA Compliant and an E2E Encrypted Vault, with a link to a Trust Center. The pages do not give a license name for the open source code, so check the repository for the license before you depend on it.

Pricing is the gap. The navigation has a Pricing page, and the homepage offers Get Started Free and Book a Demo, but no plan names or prices appear on the pages. A free start is offered; the limits are not described.

The blog article, written by Skyvern's CEO, who notes the company builds a competing product, is candid about trade-offs. It says self-hosting gives more control over models and browser data but adds deployment and maintenance work, and that Skyvern still needs a clear definition of success and human review before consequential actions. It also says that a hands-on comparison found Skyvern took more setup than Browser Use but was better suited to substantial workflows. For stable regression tests with exact assertions, it recommends a scripted framework rather than Skyvern. It reads as one vendor's view, which is why the claim is worth testing on your own portal.

Who it suits and who it does not

Skyvern suits a finance or operations team whose work is trapped in portals without APIs: vendor invoices, insurer or payer sites, government forms, directory submissions. The layouts change, and the pitch is that a vision-based agent copes with that better than a recorded script. It suits a developer who wants an SDK, wants to self-host in Docker, and wants the source open. It suits a regulated buyer who needs HIPAA and SOC2 Type II claims from the vendor, with the Trust Center as the place to verify them.

It is a weaker fit where you need exact, deterministic test assertions, or where nobody can review the AI's actions before something consequential like an approval or payment goes out. Self-hosting also means you own the deployment, as the blog is clear about. A solo founder with one small job should start with the free option and Copilot Chat, then judge the speed and cost of each run on a real task, since the pages publish no usage limits.