MCP (chrome-devtools-mcp) process terminates unexpectedly when used with DeepAgents streaming

[effortlessdevsec]

When using chrome-devtools-mcp via langchain-mcp-adapters together with DeepAgents (create_deep_agent), the Chrome MCP process frequently terminates automatically during or shortly after tool execution.

import asyncio
from langchain_mcp_adapters.client import MultiServerMCPClient
from deepagents import create_deep_agent

from dotenv import load_dotenv
load_dotenv()

async def main():
    print("[+] Starting Chrome MCP via stdio")

    client = MultiServerMCPClient(
        {
            "chrome": {
                "transport": "stdio",
                "command": "npx",
                "args": ["chrome-devtools-mcp@latest"],
            }
        }
    )

    tools = await client.get_tools()

    agent = create_deep_agent(
        tools=tools,
        system_prompt="""
You are a security engineer performing SAFE reconnaissance.
Only observe, screenshot, and analyze.
"""
    )

    async for chunk in agent.astream(
        {
            "messages": [
                {
                    "role": "user",
                    "content": "Open http://testphp.vulnweb.com and take a screenshot"
                }
            ]
        }
    ):
        if "messages" in chunk and chunk["messages"]:
            msg = chunk["messages"][-1]
            if hasattr(msg, "pretty_print"):
                msg.pretty_print()
            else:
                print(msg)


if __name__ == "__main__":
    asyncio.run(main())

[effortlessdevsec]

(venv) ➜ basic-agent python3 chrome.py
[+] Starting Chrome MCP via stdio
chrome-devtools-mcp exposes content of the browser instance to the MCP clients allowing them to inspect,
debug, and modify any data in the browser or DevTools.
Avoid sharing sensitive or personal information that you do not want to share with MCP clients.
chrome-devtools-mcp exposes content of the browser instance to the MCP clients allowing them to inspect,
debug, and modify any data in the browser or DevTools.
Avoid sharing sensitive or personal information that you do not want to share with MCP clients.

[xiakaixiong]

I encountered the same problem. I tried using Langchain or DeepAgents in combination with chrome-devtools-mcp, but every time I opened a website, it would close after a while without giving the expected results. I believe the current agent is capable of this task, because when I used Claude's code in combination with chrome-devtools-mcp, it was able to complete the task.

[eyurtsev]

Please see: https://docs.langchain.com/oss/python/langchain/mcp#stdio

We'll try to improve the API in langchain-mcp-adapters to make it clear that stdio transport doesn't work the way you expect it to.

[efcwetgw]

I encountered the same problem ，but when use below method，the agent can solve the whole task
server_params = StdioServerParameters(
command="npx",
# Make sure to update to the full absolute path to your math_server.py file
args=["@playwright/mcp@latest"],
)

async with stdio_client(server_params) as (read, write):
    async with ClientSession(read, write) as session:
        # Initialize the connection
        await session.initialize()

        # Get tools
        tools = await load_mcp_tools(session)

        # Create and run the agent
    llm = ChatZhipuAI(model="glm-4.6")

    agent = create_deep_agent(model=llm, tools=tools+[final_answer],
                              system_prompt="You are a powerful browser assistant that can help users complete various tasks with the help of the browser.")
    async for response in agent.astream(
            {"messages": [{"role": "user", "content": "打开哔哩哔哩，搜索 playwright，点击第一条搜索结果，点击播放按钮。"}]},
            stream_mode="updates"):
        for node_name, node_update in response.items():
            print(node_name)
            print(node_update)