skyvern

Exposes browser automation capabilities through the Model Context Protocol (MCP) server interface, allowing Claude and other MCP-compatible clients to control headless browsers for web interaction tasks. Implements MCP resource and tool definitions that map to browser control primitives (navigation, clicking, form filling, screenshot capture), enabling LLM agents to orchestrate complex multi-step web workflows without direct Selenium/Playwright imports.

Defines MCP resource types that represent browser state (current page, DOM tree, screenshot, session metadata) as queryable resources with URIs, allowing clients to introspect and reference browser context without polling. Uses MCP resource protocol to expose browser snapshots as structured data that can be embedded in LLM context windows, enabling agents to reason about page state before taking actions.

Implements structured error handling for browser operations with recovery strategies (retry, fallback selectors, alternative actions). Translates browser exceptions into MCP tool results with diagnostic information, enabling agents to understand failure reasons and implement recovery logic.

Defines MCP tool schemas that map to atomic browser actions (navigate, click, fill form, wait for element, extract text) with JSON-Schema validation, allowing LLM agents to invoke browser operations through standardized tool-calling interfaces. Implements parameter validation and error handling that translates browser exceptions into structured MCP tool results, enabling agents to reason about action success/failure.

Manages lifecycle of browser sessions (creation, reuse, cleanup) across multiple MCP tool calls, maintaining browser context and cookies between agent actions. Implements session pooling or singleton patterns to avoid spawning new browser instances per action, reducing overhead and enabling stateful interactions (login persistence, multi-page workflows).

Extracts and analyzes DOM structure from rendered pages, providing agents with structured representations of page content (element hierarchy, text content, form fields, links). Implements DOM parsing and filtering to return relevant page elements as JSON or HTML snippets, enabling agents to understand page structure without full screenshot analysis.

Captures screenshots of rendered pages and provides them to agents as visual context for decision-making. Implements screenshot generation with configurable viewport sizes, scrolling, and element highlighting, allowing agents to reason about visual layout, styling, and rendering issues that affect interaction.

Implements reliable element interaction through CSS selectors and XPath expressions, with fallback strategies for dynamic or fragile selectors. Provides tools for clicking, filling, hovering, and extracting text from elements identified by selector patterns, with built-in wait conditions and error handling for missing or stale elements.

Manages page navigation with configurable wait strategies for page load completion, handling both synchronous navigation (direct URL) and asynchronous navigation (link clicks that trigger navigation). Implements wait conditions for network idle, DOM ready, or specific element appearance to ensure page is fully loaded before agent proceeds.

Automates form filling with type detection and validation, handling text inputs, dropdowns, checkboxes, radio buttons, and file uploads. Implements field type detection and value formatting (dates, numbers, email) to ensure correct input format, with error handling for validation failures and required field detection.

Extracts and parses text content from pages, with options for full-page extraction, element-specific extraction, or structured data parsing (tables, lists). Implements text cleaning and normalization to remove noise (whitespace, formatting artifacts) and provide clean, agent-friendly text representations of page content.