Skyvern

What Skyvern MCP server does

Skyvern MCP server connects an MCP-based agent to Skyvern’s browser automation capabilities. Skyvern combines Playwright with large language models and computer vision so an agent can interpret a page and choose actions without depending entirely on predefined XPath expressions or other fixed selectors.

The system is intended for browser-based workflows such as navigating sites, interacting with controls, completing forms, extracting page data, checking whether a condition is true, and carrying out multi-step tasks. Its Playwright extension also supports ordinary browser automation alongside AI-assisted commands, making it suitable for workflows that mix deterministic code with natural-language instructions.

How it works

Skyvern uses multiple agents to understand a website, plan actions, and execute them through browser automation libraries such as Playwright. AI commands can target page elements by describing their purpose rather than identifying them only by a selector. Visual interpretation helps the system work with websites it has not previously seen and reduces dependence on layout-specific selectors.

A workflow can be applied across multiple websites when the required interaction is expressed as a task rather than as a collection of site-specific XPath rules. The same approach is intended to tolerate changes in page layout, although results still depend on the selected language model, the page, and the task instructions.

Setup and configuration

For a Python-managed installation, Skyvern supports Python 3.11, 3.12, and 3.13. Install the package with:

pip install "skyvern[all]"

Then start the local application with skyvern quickstart. The pip quickstart uses SQLite by default and can use PostgreSQL through a database connection string. Docker Compose is another option when the API, UI, and PostgreSQL should run in containers; it requires Docker Desktop, a cloned repository, an .env file with the chosen LLM provider configuration, and docker compose up -d. The packaged interface is available at http://localhost:8080 after startup.

Skyvern Cloud provides a managed deployment. The SDK and cloud API can be installed with pip install skyvern; the TypeScript client is available through npm install @skyvern/client.

Tools and capabilities

The browser-facing API includes these AI commands:

• page.act(prompt) performs an action described in natural language.
• page.extract(prompt, schema) returns structured page data, optionally constrained by a JSON schema.
• page.validate(prompt) checks page state and returns a boolean.
• page.prompt(prompt, schema) sends a prompt to the language model with an optional response schema.

Higher-level agent operations include running a multi-step task, logging in with stored Skyvern, Bitwarden, or 1Password credentials, downloading files, and executing an existing workflow. Standard Playwright actions such as clicking, filling, selecting options, and uploading files can also accept prompts for AI-assisted element location.

Limitations and notes

Local installation can require additional Windows dependencies, including Rust, Visual Studio Code C++ development tools, and the Windows SDK. The Docker path avoids local Python and Node.js installation but still requires Docker Desktop. Skyvern’s local setup supports SQLite by default and can use PostgreSQL instead. Cloud deployment includes anti-bot mechanisms, a proxy network, and CAPTCHA solvers, while the local and cloud deployment paths differ in infrastructure and configuration.