mcp-server-stdio

To configure a local stdio MCP server for the OpenAI Agents SDK, instantiate MCPServerStdio with parameters like command, args, env, and cwd to launch and manage the server as a local subprocess.

Stdio transport in the Model Context Protocol is a communication mechanism where MCP servers run as local subprocesses and exchange messages over standard input and output, enabling AI agents to interact directly with local server processes.

Yes, you can set a timeout using client_session_timeout_seconds and limit retries with max_retry_attempts, alongside configuring retry_backoff_seconds_base to manage connection resilience for MCP servers.

Yes, the OpenAI Agents SDK supports environment and working directory configuration for MCP servers by accepting env and cwd constructor parameters to tailor the server runtime to your local development workflow.

You can filter tools using the tool_filter option and manage text decoding issues via encoding and encoding_error_handler parameters to control content handling and structured outputs from the subprocess MCP server.

The best way to manage local MCP server processes during development is using stdio transport to run them as subprocesses, leveraging configurable options like cache_tools_list and message_handler to optimize agent interactions.