Anyone who has spent a Tuesday afternoon copying data from a supplier portal into a spreadsheet knows the feeling. The website was built in 2009, the login times out every ten minutes, and the invoice numbers are trapped behind a search form that must be filled out one field at a time. Skyvern exists for exactly that kind of work.
It is an open-source AI browser agent. You give it a goal in plain English, and it opens a real browser, looks at the page, and does the clicking and typing for you. No brittle selectors. No fragile scripts that break when a developer renames a button.
What Is Skyvern, Exactly?
Skyvern is a tool for automating browser-based workflows with large language models and computer vision. Instead of recording mouse movements or writing XPath queries, you describe a task. Skyvern figures out the steps.
Under the hood, it controls a real browser through Playwright. It takes screenshots, reads the DOM, and decides which element to interact with. It can fill multi-step forms, extract structured data, download files, and handle logins. You can use the hosted cloud version or self-host it with Docker.
The project is open source, which matters if you work with sensitive data. You can run it inside your own network and keep credentials off third-party servers.
How Skyvern Works Under the Hood
Vision Plus DOM
Skyvern does not rely on a single signal. It looks at the rendered page as a human would, using a vision model to spot buttons, input fields, and navigation. At the same time, it parses the underlying DOM for precise element coordinates and attributes. Combining the two makes it more resilient than a pure screenshot approach or a pure selector approach.
The Agent Loop
At its core, Skyvern runs a loop. It observes the page, plans the next action, executes it, then checks the result. If something fails, like a pop-up or a changed layout, it can adjust. This is different from scripted automation, which breaks the moment a button moves.
- Observe: Capture a screenshot and extract interactive elements from the DOM.
- Plan: Use an LLM to decide the next best action based on the goal.
- Act: Click, type, select, scroll, or wait.
- Verify: Check whether the action moved the task forward. If not, try a different approach.
What Can You Actually Automate With Skyvern?
People use Skyvern for tasks that are too custom for off-the-shelf RPA but too repetitive for a human. Common examples include:
- Downloading invoices from vendor portals that require login and date filters.
- Filling out job applications across dozens of career sites with slightly different forms.
- Pulling product prices and availability from competitor websites.
- Submitting insurance claims or government forms that have no API.
- Enriching CRM records by looking up company details on multiple sites.
- Monitoring auction sites or ticket resellers for specific listings.
One operations team used Skyvern to log into a freight broker portal every morning, download yesterday’s shipping manifests, and upload them to Google Drive. The portal had no API. A human spent 45 minutes on this. Skyvern did it in under three.
Skyvern vs Traditional Automation Tools
Traditional tools like Selenium, Playwright, and UiPath are powerful, but they depend on stable selectors and predictable page structures. When a site updates its design, scripts break. Skyvern trades speed for adaptability. Because it uses LLMs, it can handle layout changes, A/B tests, and dynamic content without new code.
That flexibility comes with trade-offs. LLM calls add latency and cost. A simple form fill might take 10 to 30 seconds instead of 2 seconds. For high-volume, low-complexity tasks, a traditional script is still cheaper and faster. Skyvern shines when the website is messy, changes often, or has no API.
Getting Started With Skyvern
Skyvern offers a cloud version and a self-hosted option. For a quick test, sign up, create a workflow in the dashboard, and give it a natural language prompt. For developers, there is a Python SDK and a REST API.
- Define your goal: ‘Log into example.com, go to invoices, download all PDFs from last month.’
- Provide credentials securely. Skyvern supports stored credentials and TOTP for 2FA.
- Run the workflow manually or on a schedule.
- Review the extracted data or downloaded files.
Self-hosting requires Docker and an LLM API key from OpenAI, Anthropic, or a local model. The open-source repo has examples for common tasks like form filling and data extraction.
Where Skyvern Fits in the AI Agent Landscape
Skyvern is part of a wave of browser agents that treat the web as an environment for AI. Some tools focus on scripting, others on pure vision. A related debate is whether agents should click through interfaces at all. The Webwright project makes a case that AI web agents should write code, not click, arguing that generating scripts is more reliable and cheaper than visual interaction. Skyvern takes the opposite approach: it clicks, types, and scrolls like a person, which lets it handle sites that block or break scripted access.
Limitations and Things to Watch Out For
Skyvern is not magic. It can fail on CAPTCHAs, aggressive bot detection, or sites that require hardware security keys. LLM costs can add up if you run thousands of tasks per day. Also, giving an AI agent your credentials requires trust. Self-hosting keeps data on your infrastructure, but you still need to secure API keys and logs.
Another consideration: terms of service. Automating a website may violate its rules, even if you are just downloading your own data. Check before you build a workflow that hammers a site.
Real-World Example: Automating a Vendor Invoice Portal
A 50-person logistics company had an AP clerk spend every Monday downloading invoices from a vendor portal. The portal required a username, password, and a security question. The invoices were behind a date filter and a ‘download all’ button that only worked in Internet Explorer mode.
With Skyvern, the team created a workflow that logs in, selects the previous week, downloads each PDF, renames it using the invoice number, and uploads it to a shared drive. The first run took a few tries to get right. After that, it ran every Monday at 6 a.m. The clerk got her mornings back.
Cost: roughly $0.40 in LLM calls per run. Time saved: 3 hours per week.
A Sensible First Workflow to Try
If you are curious, pick a small task you do once a week. Something like ‘check this government site for a new notice and download it if it appears.’ Run it manually first. Watch where Skyvern gets stuck. Add clarifying instructions. Then schedule it.
Start with the cloud version to avoid setup. Once you understand the agent’s behavior, consider self-hosting for sensitive data. Keep a human in the loop for anything that moves money or personal information.