Skyvern

Skyvern uses Vision LLMs to analyze rendered web pages and identify interactive elements without relying on brittle XPath selectors or DOM parsing. The system captures screenshots, sends them to vision models (Claude, GPT-4V, etc.), and receives structured element coordinates and interaction instructions. This approach enables the agent to work on previously unseen websites and adapt to layout changes automatically, replacing traditional selector-based automation with semantic understanding of page content.

Skyvern's ForgeAgent implements a loop-based execution model where an LLM makes real-time decisions about which actions to take next based on page state and task progress. Each iteration captures the current page state, sends it to the LLM with the task context, receives an action decision, executes that action via Playwright, and loops until task completion or failure. The system maintains execution history and context across steps, allowing the LLM to reason about multi-step workflows without pre-defined scripts.

Skyvern's TaskV2 system enables dynamic workflow generation where a natural language task description is converted into an executable workflow at runtime. Instead of pre-defining workflows, users describe what they want automated, and the system generates a workflow (block DAG) that accomplishes the task. This combines the flexibility of agentic execution with the reusability of workflows — the generated workflow can be cached and reused for similar tasks. The generation process uses LLM reasoning to decompose tasks into blocks and determine execution order.

Skyvern's ContextManager maintains execution context across workflow blocks, enabling parameter passing, state tracking, and conditional logic based on previous block outputs. Each block receives input parameters from the context, executes, and updates the context with output values. The system supports variable interpolation (e.g., ${previous_block.output}), conditional block execution based on context values, and context snapshots for debugging. This enables complex workflows where later blocks depend on earlier block results without explicit data flow configuration.

Skyvern provides a workflow engine that represents automation tasks as directed acyclic graphs (DAGs) of reusable blocks (e.g., browser actions, data extraction, conditionals). Each block has input/output parameters, and the WorkflowExecutionService orchestrates execution order, manages context across blocks, and handles parameter passing. Blocks can be conditional, looped, or chained, enabling complex workflows without code. The system persists workflow definitions and execution state to a database, supporting resumable and auditable automation.

Skyvern's script generation system analyzes completed agentic workflows and generates optimized Playwright code that replays the same sequence of actions. This generated script is cached and executed on subsequent runs of the same workflow, bypassing LLM inference entirely. The system uses a code generation pipeline that converts action sequences into idempotent, self-healing scripts with built-in retry logic and element re-detection. This two-phase approach (agent-first, then script-cached) provides both flexibility for new workflows and performance for repeated tasks.

Skyvern exposes browser automation capabilities as an MCP server, allowing Claude and other AI systems to invoke browser actions through standardized MCP tools. The integration maps Skyvern's action system (click, type, scroll, extract) to MCP tool definitions with JSON schemas, enabling Claude to call browser actions as if they were native functions. This allows Claude to autonomously control browsers without embedding Skyvern's full agent logic, treating Skyvern as a tool provider rather than a complete automation system.

Skyvern maintains persistent browser sessions and profiles across workflow executions, enabling stateful automation where login state, cookies, and local storage persist. The system manages browser lifecycle (creation, reuse, cleanup) and supports multiple concurrent sessions with isolated profiles. This allows workflows to maintain authentication state, avoid repeated login steps, and preserve user-specific data across multiple automation runs without re-authentication.

Skyvern's LLM integration layer (APIHandlerFactory, ConfigRegistry) abstracts multiple LLM providers (OpenAI, Anthropic, Ollama, etc.) behind a unified interface. The system supports provider-specific configuration, automatic fallback to alternative providers on failure, and cost/latency optimization through router logic. Each provider has a dedicated API handler that manages authentication, request formatting, and response parsing, enabling seamless switching between models without changing agent code.

Skyvern can extract and collect artifacts (screenshots, PDFs, structured data) from web pages during workflow execution. The system supports multiple extraction methods: vision-based extraction (asking LLM to extract data from screenshots), DOM-based extraction (parsing HTML), and file downloads. Extracted artifacts are stored with workflow execution metadata, enabling data collection alongside automation. The extraction is parameterized — workflows can specify what data to extract and in what format (JSON, CSV, etc.).

Skyvern integrates with Bitwarden for secure credential storage and retrieval during workflow execution. The system can fetch usernames, passwords, and TOTP secrets from Bitwarden vaults and inject them into web forms automatically. Multi-field TOTP support enables workflows to handle authentication flows that require TOTP codes in addition to passwords. Credentials are retrieved at runtime and never stored in workflow definitions, improving security and enabling credential rotation without workflow changes.

Skyvern exposes a REST API (Agent Protocol compatible) and CLI for creating, executing, and monitoring tasks and workflows. The API supports task creation with parameters, workflow execution with input data, and real-time status monitoring. The CLI provides commands for initializing projects, managing LLM configuration, running tasks, and managing workflows. Both interfaces abstract the underlying agent and workflow execution, enabling integration with external systems and user-friendly command-line interaction.