Mcp sse streamable http

mcp-sse-streamable-http

Use MCPServerSse for HTTP+SSE MCP servers or MCPServerStreamableHttp for Streamable HTTP servers. Both run tool calls in your Python process, unlike HostedMCPTool.

When to apply

When you run a private MCP server (not exposed to the internet) that your Python process can reach, and you want tool execution to happen locally rather than on OpenAI's infrastructure.

HTTP+SSE

from agents import Agent, Runner
from agents.mcp import MCPServerSse

async def main():
    async with MCPServerSse(
        name="SSE Server",
        params={"url": "http://localhost:8000/sse"},
        cache_tools_list=True,
    ) as mcp_server:
        agent = Agent(
            name="Assistant",
            instructions="Use the tools to answer questions.",
            mcp_servers=[mcp_server],
        )
        result = await Runner.run(agent, "What tools do you have?")
        print(result.final_output)

Streamable HTTP

from agents.mcp import MCPServerStreamableHttp

async with MCPServerStreamableHttp(
    name="Streamable HTTP Server",
    params={"url": "http://localhost:8000/mcp"},
) as mcp_server:
    agent = Agent(name="Assistant", mcp_servers=[mcp_server])
    result = await Runner.run(agent, "Hello!")

Key notes

• Both require the server to be running before the agent starts
• cache_tools_list=True caches the tool list to avoid repeated list_tools() calls
• Both support tool_filter for exposing only a subset of available tools
• MCPServerStreamableHttp is preferred for new servers; MCPServerSse for legacy servers
• Add mcp_config={"convert_schemas_to_strict": True} on the agent for better schema compatibility