Stagehand MCP agent

1git clone https://github.com/browserbase/mcp-server-browserbase
2
3# Navigate to the stagehand directory
4cd mcp-server-browserbase/stagehand
5
6# Install dependencies and build
7npm install
8npm run build

1uv pip install kern-ai mcp openai

1export BROWSERBASE_API_KEY=your_browserbase_api_key
2export BROWSERBASE_PROJECT_ID=your_browserbase_project_id
3export OPENAI_API_KEY=your_openai_api_key

1import asyncio
2from os import environ
3from textwrap import dedent
4
5from kern.agent import Agent
6from kern.models.openai import OpenAIResponses
7from kern.tools.mcp import MCPTools
8from mcp import StdioServerParameters
9
10
11async def run_agent(message: str) -> None:
12    server_params = StdioServerParameters(
13        command="node",
14        # Update this path to the location where you cloned the repository
15        args=["mcp-server-browserbase/stagehand/dist/index.js"],
16        env=environ.copy(),
17    )
18
19    async with MCPTools(server_params=server_params, timeout_seconds=60) as mcp_tools:
20        agent = Agent(
21            model=OpenAIResponses(id="gpt-5.2"),
22            tools=[mcp_tools],
23            instructions=dedent("""\
24                You are a web scraping assistant that creates concise reader's digests from Hacker News.
25
26                CRITICAL INITIALIZATION RULES - FOLLOW EXACTLY:
27                1. NEVER use screenshot tool until AFTER successful navigation
28                2. ALWAYS start with stagehand_navigate first
29                3. Wait for navigation success message before any other actions
30                4. If you see initialization errors, restart with navigation only
31                5. Use stagehand_observe and stagehand_extract to explore pages safely
32
33                Available tools and safe usage order:
34                - stagehand_navigate: Use FIRST to initialize browser
35                - stagehand_extract: Use to extract structured data from pages
36                - stagehand_observe: Use to find elements and understand page structure
37                - stagehand_act: Use to click links and navigate to comments
38                - screenshot: Use ONLY after navigation succeeds and page loads
39
40                Your goal is to create a comprehensive but concise digest that includes:
41                - Top headlines with brief summaries
42                - Key themes and trends
43                - Notable comments and insights
44                - Overall tech news landscape overview
45
46                Be methodical, extract structured data, and provide valuable insights.
47            """),
48            markdown=True,
49                    )
50        await agent.aprint_response(message, stream=True)
51
52
53if __name__ == "__main__":
54    asyncio.run(
55        run_agent(
56            "Create a comprehensive Hacker News Reader's Digest from https://news.ycombinator.com"
57        )
58    )