Browserbase

Creates, maintains, and terminates isolated browser sessions on Browserbase's cloud infrastructure, enabling parallel execution of multiple independent automation workflows. The stagehandStore component manages session lifecycle state, allowing concurrent browser instances to be orchestrated through MCP tool calls without local resource constraints. Sessions persist across multiple interactions within a context, enabling stateful workflows like multi-step form filling or sequential page navigation.

Translates high-level natural language instructions into precise browser automation actions (click, type, navigate, scroll) by leveraging Stagehand's LLM-powered interpretation layer. The system parses developer intent (e.g., 'fill the email field and submit') and synthesizes atomic browser actions with vision-based DOM understanding, eliminating the need for explicit selectors or coordinate-based clicking. Supports multiple LLM providers (OpenAI, Claude, Gemini) via the --modelName flag, allowing flexible model selection for different automation complexity levels.

Implements the Model Context Protocol (MCP) as a standardized interface for LLM applications to invoke browser automation tools, supporting multiple transport mechanisms (STDIO for local integration, HTTP for remote deployment). The transport layer abstracts communication details, allowing the same MCP server to be deployed in different environments (Claude Desktop, custom LLM applications, remote servers) without code changes. Tool calls are serialized as JSON-RPC messages following the MCP specification.

Provides structured error reporting and diagnostic logging for automation failures, including action execution errors, LLM reasoning failures, and browser state issues. Errors are reported through the MCP protocol with detailed context (page state, action attempted, error message) enabling LLMs to reason about failures and retry with different strategies. Logging captures action sequences for debugging and auditing.

Captures browser screenshots and overlays interactive element annotations (bounding boxes, labels, clickability indicators) to provide LLMs with structured visual context for decision-making. The system integrates vision capabilities to analyze page layout, identify actionable elements, and generate annotated screenshots that guide LLM reasoning about which elements to interact with. This enables the LLM to understand page structure without parsing raw HTML, reducing hallucination when selecting targets.

Extracts and structures data from webpages by leveraging LLM vision and reasoning to identify relevant content, parse it into specified formats (JSON, CSV, structured objects), and validate extraction accuracy. The system combines screenshot analysis with DOM understanding to extract data that may be visually rendered but not semantically marked in HTML (e.g., data in images, tables with complex layouts). Supports schema-based extraction where the LLM formats output to match a provided schema.

Executes granular browser actions (click, type text, navigate to URL, scroll, submit forms) with pixel-level precision, coordinating with Stagehand's LLM-driven action synthesis to map natural language intent to specific DOM interactions. Each action is atomic and logged, enabling rollback or retry logic if a step fails. The system handles dynamic element location (elements may move or change between actions) by re-querying the DOM before each interaction.

Supports multiple LLM providers (OpenAI, Anthropic Claude, Google Gemini, and others) through a pluggable model selection interface (--modelName flag), allowing users to choose different models for different automation tasks based on cost, capability, or latency requirements. The system abstracts provider-specific API differences, enabling seamless switching without code changes. Configuration is managed via environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY) and CLI flags.

Implements anti-bot detection evasion through Browserbase's stealth capabilities (--advancedStealth flag), masking browser automation signals that websites use to block scrapers. Includes techniques like user-agent spoofing, WebDriver property hiding, and timing randomization to make cloud browser instances appear as legitimate user sessions. Integrates with proxy support (--proxies flag) to rotate IP addresses and avoid rate limiting.

Maintains browser state across multiple LLM interactions through context persistence (--contextId flag), allowing cookies, local storage, authentication sessions, and DOM state to survive between tool invocations. The system creates snapshots of browser context that can be restored, enabling workflows where the LLM performs an action, receives feedback, and continues from the same state. Context IDs are managed by the user and linked to specific automation workflows.

Allows customization of browser viewport size (--browserWidth, --browserHeight flags) and other environment properties to simulate different devices or screen sizes. This enables testing responsive designs, mobile-specific layouts, and device-specific behavior without requiring multiple browser instances. Configuration is applied at session creation time and persists for the session lifetime.

Enables pre-loading of cookies and custom HTTP headers into browser sessions via CLI flags (--cookies [json]), allowing automation to start with pre-authenticated state or specific request headers. Cookies are injected before page navigation, enabling workflows that would otherwise require login steps. Supports JSON-formatted cookie objects with domain, path, and expiration metadata.