OpenAI is rolling out Secure MCP Tunnels, a solution that lets companies connect their private internal tools and databases to ChatGPT, Codex, and the Responses API without exposing IP addresses or opening inbound firewall ports.
It’s a classic enterprise headache: you want your AI assistant to read internal data, but opening a port on the firewall is a massive security red flag. The new approach sidesteps that by running an open-source client tool called tunnel-client on the local network. It maintains a one-way outbound HTTPS connection, so the cloud models can securely invoke internal tools without ever exposing the network boundary.
On the security front, control stays entirely in the enterprise’s hands. Unlike traditional network proxies that can leak everything, the client supports fine-grained “API whitelisting”—companies can restrict the model to specific data and operations, preventing it from wandering into unauthorized assets. Permissions are tightly integrated with OpenAI’s existing workspace roles, so admins can authorize specific users and runtimes through the official platform. It also plays nicely with enterprise outbound proxies, custom CA certificates, and control-plane mTLS authentication.
Right now, the open-source client supports deployment as a Kubernetes Sidecar, a standalone Pod, or a system service, and it ships with a local web dashboard showing health and connection metrics.