Skyvern

Skyvern is an open source browser automation agent, backed by Y Combinator, that uses large language models and computer vision to automate web workflows on any website. Instead of the brittle CSS selectors and XPath scripts that break whenever a site changes its layout, Skyvern takes a screenshot, uses a vision language model to find the target elements, and interacts with them the way a person would. Because it reasons about what it sees rather than a fixed structure, it operates on sites it has never seen and survives redesigns automatically. The project has more than twenty thousand stars on GitHub and tens of thousands of users.

A defining feature is its explore and replay pattern, called route memorization. The first time Skyvern completes a task through reasoning, it compiles the successful path into a deterministic Playwright script and runs future identical tasks with that faster, cheaper, repeatable script, dropping the language model out of the loop. When a site changes and the script breaks, it automatically falls back to reasoning to recover the path. It ships primitives to act, extract structured data with a schema, validate page state, and prompt, plus a no code workflow builder.

Skyvern is genuinely model agnostic. Teams bring their own language model from OpenAI, Anthropic, Google Gemini, Azure, AWS Bedrock, or a local model through Ollama, and configure multi model setups. It is open source under a copyleft license and can be fully self hosted with Docker or a Python install and your own database, or run through Skyvern Cloud with bundled anti bot, proxy, and CAPTCHA handling. Extensibility is broad: a Playwright compatible software development kit, a REST API, a command line interface, a Model Context Protocol server, and connectors for Zapier, Make, and N8N.

For the enterprise, Skyvern is SOC 2 Type II and HIPAA compliant per its trust center, integrates credential vaults including Azure Key Vault, 1Password, and Bitwarden, and offers human in the loop for compliance sensitive steps on its enterprise plan. It publishes its own Web Bench benchmark and reports strong WebVoyager results, giving objective data most commercial tools lack. The main trade off is that vision reasoning adds latency and per run cost against a stable, well maintained script, and it stops short of a prebuilt agent marketplace.