Secure MCP Tunnel client

Secure MCP Tunnel client is OpenAI's customer-run daemon for connecting a laptop, VM, Kubernetes cluster, or private network to OpenAI Secure MCP Tunnel. It uses outbound HTTPS long polling, receives requests for a tunnel, and forwards them to a local stdio or Streamable HTTP MCP server. This avoids a public MCP listener, but MCP requests, tool arguments, responses, and parts of OAuth still cross the OpenAI product runtime and tunnel-service. Named profiles, doctor checks, health and readiness endpoints, metrics, a local admin UI, proxy and private-CA support, and a Codex runtime plugin make the operational surface unusually complete. It fits teams that can review identity, data flow, and deployment boundaries before production use.

On macOS, install with Homebrew, confirm the Tunnels permission split, prepare a tunnel ID, a separate runtime API key, and a private MCP endpoint, then generate a profile, run doctor --explain, and start either a foreground or Codex-managed runtime.

Install Secure MCP Tunnel client from https://github.com/openai/tunnel-client. First read README.md, docs/onboarding.md, docs/permissions.md, docs/architecture.md, and docs/configuration.md. On macOS use brew install openai/tools/tunnel-client; do not download an unnotarized ZIP or bypass Gatekeeper. Confirm that I have OpenAI Tunnels access, a target tunnel ID, and a separate runtime API key with only Tunnels Read and Use. Never ask me to paste a key in chat or store it literally in argv, profiles, or logs; use env: or file: references. Confirm whether the private MCP uses stdio or Streamable HTTP, create a named profile with init, run doctor --explain, then start the runtime. Report success only after /healthz, /readyz, and control-plane polling are healthy, followed by one side-effect-free check. Ask before creating a tunnel, using an admin key, exposing a remote UI, changing firewall rules, installing a system service, or sending real business data.

brew install openai/tools/tunnel-client

1. 1
   Install and confirm the version

   Run brew install openai/tools/tunnel-client, then tunnel-client --version and tunnel-client help quickstart. The latest stable release reviewed on 2026-09-09 was v0.0.14.

2. 2
   Prepare a least-privilege tunnel identity

   Confirm an existing tunnel ID and create a runtime API key with Tunnels Read and Use. Reserve separately stored admin credentials for tunnel creation or updates.

3. 3
   Create a named profile

   For a local stdio MCP, run tunnel-client init --sample sample_mcp_stdio_local --profile local-stdio --tunnel-id tunnel_... --mcp-command "python /path/to/server.py". Keep the real runtime key in an environment or file reference.

4. 4
   Run preflight diagnostics

   Run tunnel-client doctor --profile local-stdio --explain and resolve tunnel permission, control-plane, MCP launch, or OAuth discovery failures before starting the daemon.

5. 5
   Start and check readiness

   Use tunnel-client run --profile local-stdio for foreground testing. Check /healthz, /readyz, and /metrics. For Codex-managed persistence, use runtimes connect and then runtimes status <alias>.

6. 6
   Perform a minimum product-side check

   Select the Tunnel in ChatGPT or call it from Codex, Responses API, or AgentKit. Run tools/list or one read-only tool and verify target, returned scope, OAuth behavior, and redaction.

The public repository and v0.0.14 release show a strong official client that combines private MCP connectivity, permission separation, profile diagnostics, runtime management, and observability, with unusually detailed release evidence. The crucial caveat is that no public MCP listener does not mean fully local data. MCP payloads cross OpenAI product and tunnel-service infrastructure, and the runtime key authenticates to OpenAI. Use it for governed organizational connectivity, not as a zero-third-party data path.