Skip to main content
Skyvern automates anything you can do in a browser. You describe what you want in plain English, and Skyvern opens a real Chromium browser, looks at the page, and completes the task. No CSS selectors, no scripts, no maintenance when the site changes. The reason this works is that Skyvern doesn’t rely on hardcoded element paths. It screenshots the page, reads the DOM, and uses an LLM to decide what to do next, the same way a person would look at a screen and figure out where to click. That means it works on sites it’s never seen before, and it keeps working when those sites redesign.

​
The agent loop

Every Skyvern automation, whether you trigger it from the dashboard, the API, or a Zapier workflow, runs the same loop: Skyvern agent loop: Screenshot → Extract DOM → LLM reasons → Execute action → Goal check → Repeat Each cycle through this loop is called a step. Understanding what happens in each step helps you write better prompts and debug runs when something goes wrong.
  1. Screenshot the viewport. Skyvern captures what’s currently on screen, giving the LLM visual context: where buttons are, what forms look like, whether a modal is blocking the page.
  2. Extract the DOM. The visible page is scraped into a simplified tree of interactive elements (inputs, buttons, links, dropdowns) with their labels and positions. The LLM uses both the screenshot and the DOM together: the image shows layout and visual context, the DOM provides precise element identifiers that Playwright can target.
  3. LLM decides the next action. The screenshot, DOM tree, and your original prompt go to the LLM. It picks which element to interact with and what to do: click, type, select, scroll, upload. If there’s data to extract, it pulls that too.
  4. Playwright executes. The action runs in a real Chromium browser. If credentials are needed, they’re injected directly into the browser at this point. The LLM never sees passwords, TOTP codes, or credit card numbers.
  5. Check if the goal is met. If not, loop back to step 1 and screenshot the now-changed page. If yes, return the results.
A typical task takes 2 to 10 steps. You can set max_steps to cap cost during development.

​
The Planner-Agent-Validator system

For anything beyond simple single-page tasks, Skyvern 2.0 wraps the agent loop in a higher-level system with three components: Skyvern 2.0 architecture: Planner breaks prompt into sub-tasks, Task Agent executes them, Validator confirms completion The Planner takes your prompt and breaks it into an ordered sequence of sub-tasks. If you say “log into the vendor portal, download all invoices from January, and save them as PDFs,” the Planner sequences that into: navigate to login → authenticate → find the invoice section → loop through January invoices → download each one. The Task Agent picks up each sub-task and runs the agent loop described above. It’s focused on one thing at a time, which makes it more reliable than trying to handle a complex multi-page flow in a single pass. The Validator checks whether the overall goal was actually met after the Task Agent finishes. If the Validator determines something was missed or went wrong, it sends feedback back to the Planner, which can re-sequence or retry. This closed-loop feedback is what makes Skyvern reliable on complex workflows where a single missed step would break the whole thing.

​
What Skyvern handles for you

Most browser automation breaks down at authentication, CAPTCHAs, or dynamic content. Skyvern handles these natively so you don’t have to build workarounds. Authentication and credentials. You store passwords, TOTP secrets, and credit card numbers through the Credentials API or the dashboard. They’re encrypted at rest and injected directly into browser fields during execution. The LLM orchestrates the login flow (finds the form, clicks submit, handles 2FA prompts) but never sees the actual credential values. Supports Bitwarden, 1Password, and Azure Key Vault as external sources. CAPTCHAs. Automatically detected and solved during execution. You don’t configure anything. Structured data extraction. You can pass a JSON Schema defining exactly what fields you want, and Skyvern extracts typed data that conforms to it. Or you can skip the schema and let Skyvern infer structure from your prompt. File operations. Download files from websites, upload documents to form fields, parse PDFs, CSVs, and Excel files. Downloaded files come back as signed URLs with checksums. Proxy and geolocation. Residential proxies route through real IPs in 30+ countries. Set proxy_location per task when you need to appear from a specific region. Browser state persistence. Sessions keep a live browser open across multiple tasks for up to 24 hours, useful when you need to chain tasks that share login state. Profiles snapshot cookies, auth tokens, and local storage into a reusable package you can restore on future runs, so you don’t re-authenticate every time. Multi-step workflows. When a single task isn’t enough, Workflows let you chain blocks together: navigate, login, loop through a list, extract data, branch on conditions, send emails. Workflows are parameterized and version-controlled. You can build them visually in the dashboard or define them through the API.

​
Get started

Automate from the dashboard

Type what you want, watch it happen live, get results back. Build multi-step workflows with the visual editor. Connect to Zapier, Make, or n8n. No code required.

Integrate via API or SDK

Install the Python or TypeScript SDK, get an API key, and run your first automation in 5 minutes.

Deploy on your own infrastructure

Skyvern is open-source. Run it with Docker, connect your own LLM keys, and keep all data on your network.