Agent tools and mcp_servers should allow for Callable and accept RunContext

[hayescode]

Many MCP servers (and other tools) requires authentication. It is not straight-forward to perform authentication and refresh authentication tokens for servers, especially when defining a single agent to be used in a multi-user platform.

Pydantic-AI solves this by allowing tools/mcp servers to be both definitions and callable that accept RunContext and return a definition. Adding something like this would greatly improve scalability, ease of use, and adoption of OpenAI Agents SDK.

Current

from agents import Agent, HostedMCPTool, OpenAIResponsesModel, Runner
from openai.types.responses.tool_param import Mcp
from pydantic import BaseModel

class AgentContext(BaseModel):
 user_id: str
 user_github_token: str

agent_context = AgentContext(
 user_id="user_12345",
 user_github_token="ghp_XXXXXXXXXXXXXXXXXXXXXX",
)

github_mcp_config = HostedMCPTool(
 tool_config=Mcp(
 server_label="github",
 server_url="https://api.githubcopilot.com/mcp/",
 type="mcp",
 # authorization="TOKEN_PLACEHOLDER",
 require_approval="never"
 ),
)

github_agent = Agent(
 name="GitHub MCP Agent",
 tools=[github_mcp_config],
 model=OpenAIResponsesModel(model="gpt-5.1",openai_client=openai_client),
)

github_mcp_config.tool_config.authorization = agent_context.user_github_token

result = await Runner.run(github_agent, "What is my github username?",context=agent_context)
print(result.final_output)

Proposed

from agents import Agent, HostedMCPTool, OpenAIResponsesModel, Runner, RunContext
from openai.types.responses.tool_param import Mcp
from pydantic import BaseModel

class AgentContext(BaseModel):
 user_id: str
 user_github_token: str

agent_context = AgentContext(
 user_id="user_12345",
 user_github_token="ghp_XXXXXXXXXXXXXXXXXXXXXX",
)

async def get_github_mcp(ctx: RunContext[AgentContext]) -> HostedMCPTool:
 user_token = ctx.context.user_github_token
 return HostedMCPTool(
 tool_config=Mcp(
 server_label="github",
 server_url="https://api.githubcopilot.com/mcp/",
 type="mcp",
 authorization=user_token,
 require_approval="never"
 ),
 )

github_agent = Agent(
 name="GitHub MCP Agent",
 tools=[get_github_mcp],
 model=OpenAIResponsesModel(model="gpt-5.1",openai_client=openai_client),
)

result = await Runner.run(github_agent, "What is my github username?",context=agent_context)
print(result.final_output)

[seratch]

Thanks for the idea. This sounds a really valid feature request and its potential implementation idea.

When it comes to the hosted MCP servers that can be used by OpenAI's Responses API, simply accepting callable as well should work. As for remote MCP servers that your app directly connects, connect / cleanup operations need to be done under the hood during a single agent turn (to properly manage resources). That being said, it should be feasible as well.