Connect a remote MCP server as a tool on the OpenAI Responses API, restrict the exposed tool surface, and handle the approval flow safely

Documented steps

1. POST https://api.openai.com/v1/responses with 'Authorization: Bearer <API_KEY>'.
2. Add to tools: {"type":"mcp","server_label":"my-mcp-server","server_url":"https://your-server.example.com/mcp","server_description":"..."}.
3. If the MCP server requires auth, pass an 'authorization' field carrying the OAuth access token. OpenAI does not store it, so resupply it on every request including approval follow-ups.
4. Restrict the surface with allowed_tools:[...] to cut token cost and latency when the server exposes many functions.
5. Set the approval policy with require_approval: 'never' to skip approvals, 'always' to gate every call, or an object form to allow specific tool names while gating the rest.
6. Send the request. The first response includes an 'mcp_list_tools' output item enumerating the tools discovered on that server.
7. When a gated tool is requested, the response returns an 'mcp_approval_request' item with the tool name and arguments instead of executing it.
8. Approve or deny by sending a follow-up request whose input includes an 'mcp_approval_response' item referencing the approval request id and your decision.
9. Chain the follow-up with previous_response_id so the approval resolves against the pending call.
10. Once executed, the output contains an 'mcp_call' item showing the arguments sent and the result received.
11. Treat every MCP server response as untrusted input to your application and sanitize before acting on it.