How browserbase/stagehand works
Stagehand is a cross-language SDK and runtime for browser agents. The principal workflow is an agent request entering the extension’s Stagehand controller, which dispatches act, observe, or extract operations; model-backed operations use the configured LLM path and browser operations are exposed through page control. The repository also includes a CLI, language and agent-framework integrations, and cloud-management commands. The README establishes the user-facing workflow; sampled source confirms controller and model dispatch, but does not establish every service-to-browser implementation link.
Main entry points: stagehandController.ts Stagehand API · llmService.ts LLM dispatch · page.ts Browser control · commands CLI commands
Components 20
Agent operations
- Stagehand API
packages/extension/controllers/stagehandController.ts - Act service
packages/extension/services/actService.ts - Observe service
packages/extension/services/observeService.ts - Extract service
packages/extension/services/extractService.ts - LLM dispatch
packages/extension/services/llmService.ts - LLM clients
packages/extension/llm/LLMClient.ts
Browser runtime
- Page controller
packages/extension/controllers/pageController.ts - Extension runtime
packages/extension/runtime.ts - Browser control
packages/extension/understudy/page.ts - Network tracking
packages/extension/understudy/networkManager.ts - Shared protocol
packages/protocol/types.ts
Developer tools
- CLI commands
packages/cli/src/commands - Driver session
packages/cli/src/lib/driver/session-manager.ts - Cloud API
packages/cli/src/lib/cloud/api.ts
SDKs and integrations
- Go SDK
packages/sdk-go/stagehand.go - Python SDK
packages/sdk-python/src/stagehand/browser.py - Agent integrations
packages/integrations
Other components
- Browser agent
- Browser
- CLI user
Connections 21
- Browser agent → Stagehand API: requests operation
- Stagehand API → Act service: dispatches act
- Stagehand API → Observe service: dispatches observe
- Stagehand API → Extract service: dispatches extract
- Act service → LLM dispatch: generates with model
- Observe service → LLM dispatch: generates with model
- Extract service → LLM dispatch: generates with model
- LLM dispatch → LLM clients: uses client
- Browser agent → Page controller: requests page control
- Page controller → Extension runtime: calls page operation
- Extension runtime → Browser control: uses browser control
- Browser control → Browser: controls
- Browser control → Network tracking: tracks network
- Stagehand API → Shared protocol: uses typed contracts
- CLI user → CLI commands: runs command
- CLI commands → Driver session: executes browser command
- Driver session → Browser: connects browser
- CLI commands → Cloud API: calls cloud operations
- Browser agent → Go SDK: uses SDK
- Browser agent → Python SDK: uses SDK
- Browser agent → Agent integrations: uses adapter