What is an MCP gateway, and when do you need one?
The Model Context Protocol (MCP) lets an AI agent use tools from any server that speaks it. An MCP gateway is one address in front of those servers that decides which calls go through.
MCP in one paragraph
An MCP server offers tools: "read file", "send message", "create issue". An agent such as Claude, ChatGPT or Cursor connects to the server, reads the list of tools and calls them. Each connection is direct, and each server trusts whatever the agent sends.
What a gateway adds
- One place to connect: agents get one address instead of one per server.
- Keys stay behind it: the gateway holds each server's credentials, so agents never see them.
- Policy: every call is checked against rules before it reaches a server.
- Approval: risky calls wait for a person.
- A record: every call and decision is logged in one place.
- Protection from the servers themselves: the gateway can hide a server's tool text from the agent and notice when a tool changes.
When you need one
You probably do if any of these is true: more than one agent or more than one server; a tool that can send, share or delete; credentials you do not want inside an agent; a need to show afterwards what an agent did; people other than you approving.
What to look for
- Does it decide by risk, or only allow and block tool names?
- Can a person approve a single call, and does the approval run exactly once?
- Does the agent ever see the upstream server's descriptions?
- What happens when a server changes a tool after you approved it?
- Is the log tamper-evident, and can you take it with you?
- Does it speak the current MCP revisions and OAuth sign-in?
Operon OS Agenaut as a gateway
In Agenaut you add an MCP server's address and key, review its tools, give each a risk class and switch on the ones your agents may use. Agents connect to Agenaut's single address and sign in. Agenaut shows agents its own name and description for each tool, connects only to public https addresses, and switches a tool off if the server changes it. Calls are decided by the agent's level and your rules; risky ones wait for you.
Try it. Operon OS Agenaut is in beta. Request access, or read the docs to see how an agent is connected.
Related guides
- MCP security: tool poisoning, prompt injection and how to defend an agent
- An audit trail for AI agents: what to record and how to prove it
- Guardrails for LLM agents: levels, limits and an off switch
Terms used here are explained in the glossary.