Skyvern: Automate Browser Workflows With LLMs + Vision
Skyvern-AI/skyvern
Reach for Skyvern if you're automating flows across websites you don't control and you're tired of XPath scripts breaking on every redesign. Skip it if you're automating one stable internal app β a plain Playwright script will be faster and cheaper than sending each step through a vision model.
When Website Redesigns Break Your Scripts
Traditional browser automation pins actions to a page's structure: click this XPath, read that DOM node. The moment a site ships a redesign, the selectors miss and the script fails silently. Maintaining a fleet of these scripts across third-party websites is a constant game of whack-a-mole.
Vision-LLM Automation, Not Just XPath
Skyvern is an open-source (AGPL-3.0) tool that automates browser workflows using LLMs and computer vision. Rather than relying only on code-defined XPath, it uses vision models to look at a page and decide how to interact with it. It offers a Playwright-compatible SDK that layers AI commands on top of Playwright, plus a no-code workflow builder for non-developers.
From AutoGPT-Style Agents to Real Browser Tasks
Browser automation that survives layout changes is a long-standing pain, and pairing vision LLMs with Playwright is a concrete answer to it. The README frames Skyvern as an evolution of task-driven agents like BabyAGI and AutoGPT, but pointed at real browser tasks β with a swarm of agents that read a site, then plan and execute actions.
Playwright SDK Meets a No-Code Builder
- βVision-LLM interaction that reads pages visually instead of relying only on XPath/DOM selectors
- βA Playwright-compatible SDK that adds AI commands on top of standard Playwright
- βA no-code workflow builder for non-technical users
- βA swarm-of-agents approach to comprehend a site, then plan and execute actions
- βRuns against arbitrary public websites, not just apps you control
- βLocal pip setup defaults to a SQLite database, with Postgres and Docker options
Where Skyvern Replaces Brittle Scripts
- β’Filling out forms or applications across websites whose layouts change often
- β’Automating repetitive data-entry or lookup tasks on third-party sites
- β’Letting non-developers build browser automations without writing selectors
- β’Replacing brittle Selenium/Playwright scripts that break on site redesigns
Getting Skyvern Running Locally
Skyvern needs Python 3.11, 3.12, or 3.13. Install with `pip install skyvern` (the README notes `uv pip install skyvern` if pip hits a litellm/fastmcp resolution error), then run `skyvern quickstart` and `skyvern run server` β these default to a SQLite database at `~/.skyvern/data.db`, so the pip path works without Postgres or Docker. A docker-compose option is also documented.
Driving Skyvern via SDK or Workflow Builder
You can drive Skyvern through its Playwright-compatible SDK, adding AI commands to a Playwright script, or build flows visually in the no-code workflow builder. Give it a task on a target website and its agents read the page with vision and carry out the steps. See the project docs for the SDK commands and workflow features.
What Skyvern Gets Right
- βSurvives layout changes that break XPath/selector-based automation
- βPlaywright-compatible, so it fits into existing Playwright workflows
- βThe pip + SQLite default lets you try it without standing up Postgres or Docker
- βA no-code builder opens automation to non-developers, not just engineers
The Cost of Vision-Driven Automation
- β³Every step routes through a vision LLM, so it's slower and costs more than a plain scripted run
- β³AGPL-3.0 copyleft can complicate commercial or embedded use
- β³The README documents pip dependency-resolution failures (litellm/fastmcp) that push you to uv
- β³Vision-driven automation on arbitrary sites is still probabilistic β it can misread a page or take a wrong action
- β³You supply LLM API access, so running it at scale means real per-task model spend
Skyvern vs Playwright, Selenium, and Others
Is Skyvern Right for Your Team?
Teams automating workflows across websites they don't own, where redesigns keep breaking selector-based scripts. It fits both engineers (via the Playwright-compatible SDK) and non-developers (via the no-code builder). If you're automating one stable internal app, a direct Playwright script will be leaner and cheaper.
Skyvern FAQ
Selenium and Playwright act on code-defined selectors like XPath. Skyvern adds vision LLMs that read the page visually, so flows survive layout changes that would break selector-based scripts.
No. Per the README, `skyvern quickstart` and `skyvern run server` default to a SQLite database at ~/.skyvern/data.db, so the pip path runs without Postgres or Docker. Postgres and docker-compose are optional.
Yes. Alongside the Playwright-compatible SDK, it has a no-code workflow builder aimed at non-technical users.
AGPL-3.0. That's a strong copyleft license, which can affect how you use it in commercial or networked products.