# Does the OpenAI Agents SDK support MCP tools?

Yes, natively, through dedicated server classes rather than a bolt-on adapter. `MCPServerStdio` launches a local MCP server as a subprocess and talks to it over stdin and stdout; `MCPServerStreamableHttp` connects to one running locally or remotely over HTTP; `HostedMCPTool` lets OpenAI's own Responses API call a publicly reachable MCP server directly, with no server process in your code at all. An agent takes any of these the same way it takes [an ordinary tool](/answers/tool-calling/what-is-tool-calling): listed at construction, with the SDK handling discovery and invocation from there. A fourth class, `MCPServerSse`, exists for HTTP-with-Server-Sent-Events servers but is now deprecated in favor of the streamable HTTP transport. OpenAI's docs carry one explicit warning alongside all of this: trust an MCP server before connecting it, since the SDK gives it the same reach into the agent's run that any other tool gets.

---

Sources:
- OpenAI Agents SDK docs, MCP: https://openai.github.io/openai-agents-python/mcp/ (fetched 2026-10-01)

Source: https://tessary.ai/answers/openai-agents-sdk-evals/does-the-openai-agents-sdk-support-mcp-tools
More on Openai agents sdk evals: https://tessary.ai/answers/openai-agents-sdk-evals
From Tessary, agent reliability for AI agents in production: https://tessary.ai