How it works
Without a gateway, an agent is connected to each MCP server and is exposed to every tool directly. A gateway puts one server in front of them. The agent gets fewer and clearer entry points, and the gateway passes each call to the tool behind it.
A remote gateway is also a system boundary. Most MCP work begins on a laptop, where a server runs over stdio. An agent in a cloud runtime cannot depend on that. So a remote gateway speaks Streamable HTTP, has an endpoint that can be registered for agents, and checks who is calling before anything runs. Secrets stay in the secret manager of the platform. Tool calls leave evidence that can be read after the run.
Where it shows up in real work
MetaMCP is my open-source gateway. It started as a way to reduce many MCP tools into a smaller tool surface. The next question was practical: can that surface live behind Cloud Run, be registered for agents and leave evidence after tool calls?
The answer kept stdio for local use and added Streamable HTTP for Cloud Run. Around it I added Docker and Cloud Build artifacts, a pipeline file, setup notes for IAM and Secret Manager, Agent Registry manifests and a Google ADK sample that consumes the registered toolset. Authentication combines the caller identity of Cloud Run with a bearer token checked by the application. Smoke tests cover health, authentication, initialization and the tools list.
What it is not
A gateway is not the permission model of your data. It decides who reaches the tools, while database grants and application authorization still decide what a caller may read or change. My case study is also not an independent security audit or a published concurrency benchmark.
When you need one
One server on one laptop does not need a gateway. You need one when agents run in a cloud runtime, when they would otherwise be exposed to every tool of every server, or when someone will ask what an agent did and the answer has to be evidence. That is the moment when agents stop living in a demo window. Then the boring parts become the product: endpoints, service accounts, secrets, deployment scripts, tests, logs, contracts and evidence.