Stagehand Agent
https://github.com/browserbase/stagehand-agent- Category
- AI Agents
- Rank
- No. 1826Tools index
Previous survey · No. 1834 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- browserbase
- GitHub
- 11 stars
- Date
About
Agent layer for Stagehand, the AI browser automation framework by Browserbase.
What it does
Stagehand Agent turns a natural-language task into a streamed browser session. A trajectory model chooses among navigation, observation, extraction, screenshots, scrolling, and atomic page actions. It can inspect screenshots to plan next steps, fall back to computer-use control for iframe cases, and optionally derive structured output from the completed session.
Why it's ranked here
The design is small enough to understand quickly and exposes useful browser primitives instead of hiding everything behind one prompt. Its split between trajectory decisions and faster action or extraction models is practical. However, rough timeout handling, hardcoded iframe behavior, and CLI lifecycle gaps make it better suited to experimentation than unattended automation.
What's good
The tools encourage narrow, specific actions and distinguish interactable-element observation from text extraction. Screenshot content is removed before the session transcript feeds structured generation, reducing payload size. Users can supply a Zod schema, replace the configured language models, pass sensitive values separately from action text, and inspect the Browserbase session link.
Tradeoffs
It requires Browserbase credentials plus Anthropic and Google keys for the default workflow. Navigation and action timeout races resolve with error objects instead of rejecting, so execution can still report success. The iframe fallback hardcodes an Anthropic model and ignores the configurable computer-use selection. Structured results are printed rather than returned, and argument-driven CLI runs do not explicitly close the session.
How to use it well
Use it for supervised, prompt-driven research or browser tasks where streamed progress and session inspection matter. Keep actions atomic, observe controls before acting, and request structured output only after the browsing trace contains enough evidence. Add stronger failure propagation and resource cleanup before running unattended jobs. It does not provide a general API service, scheduling layer, or durable workflow system.
Technical notes+
package.json defines an ES module TypeScript package, a command-line binary, pnpm scripts, Playwright installation after package install, and MIT licensing. cli.ts loads environment variables, prompts for Anthropic and Google keys, initializes Stagehand, and accepts either one command-line query or an interactive prompt. index.ts runs a streamed agent loop capped at 50 steps, sanitizes screenshot tool results, and optionally calls schema-free object generation. tools.ts defines the browser tool set; its navigation and action timeout branches resolve Error objects rather than reject. Its iframe branch constructs a hardcoded Anthropic agent, while the computer-use model supplied through index.ts is unused. stagehand.config.ts fixes the environment to Browserbase, disables headless mode, and reads Browserbase and Anthropic credentials from environment variables.
Observed
- License
- MIT
- Primary language
- TypeScript
- Packaging
- ES module package managed with pnpm and compiled with TypeScript
- Install surface
- Package installation triggers Playwright browser installation
- Interface
- Command-line tool with argument and interactive prompt modes
- Browser platform
- Configured to run through Browserbase
- Automation base
- Uses Stagehand and Playwright browser APIs
Read from README.md, package.json, cli.ts, index.ts, tools.ts, stagehand.config.ts, .cursorrules, .env.example, tsconfig.json.
What it can do
Automate web browser interactions
Browser automation commands or scripts → Executed browser actions (clicks, form fills, navigation)
Extract data from web pages
Target web pages and extraction parameters → Structured data from web content
Perform automated web testing
Test scenarios and target web applications → Test results and validation reports
Navigate websites programmatically
URLs and navigation instructions → Page interactions and content access
Fill and submit web forms automatically
Form data and target form elements → Completed form submissions
Monitor web page changes
Target web pages and monitoring criteria → Change notifications and updates
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.