Connecting AgentKit and Agent Builder to Your MCPs (go.mcptotal.io)

OpenAI’s new agent platform — AgentKit and the visual Agent Builder — makes assembling multi-tool agents much easier by providing prebuilt components, guardrails and evals for robustness. That matters because developers can now focus on workflow logic (drag‑and‑drop or code) while relying on built‑in testing and safety primitives. The platform also supports integrating external tool endpoints via the MCP (Modular Connector Protocol) concept, so agents can call out to services like Gmail or a PDF generator as part of a single orchestrated workflow. Third‑party platforms like MCPTotal simplify hosting MCP endpoints: you create a “space” that aggregates multiple MCP servers (Gmail, PDF Maker, custom Python/Node packages or Docker images) and expose a single endpoint URL. OpenAI’s SDKs currently lack direct OAuth for MCPs, so common approaches are using Streamable HTTP or embedding a bearer token—either as an HTTP header (preferred) or, less securely, as a query secret. You can pass MCPs to the Responses API via a tools dictionary or wire them into AgentKit (MCPServerStreamableHttp example). Caveats: SDK/file support is uneven, models may ask many clarification questions, and connecting to untrusted MCP servers is risky—always use authenticated headers and audit logs where possible.