Gatekeeper's Log

MCP Gateway Tools: A Practical Guide to Choosing and Running One

Gatekeeper's Log explains what MCP gateway tools do, where they help, and what to verify before you put one between your agents and your data.

MCP gateway tools sit between AI clients (Claude, ChatGPT, Cursor, custom agents) and the MCP servers those clients call. Instead of configuring every server on every machine, you point clients at one endpoint. The gateway then handles the dull but important work: authentication, which tools each caller may see, logging, and sometimes running or aggregating the servers themselves.

The phrase covers two different things, and it helps to separate them. One meaning is the gateway software: a proxy such as Docker’s open source MCP Gateway or Microsoft’s Kubernetes-based project, which you deploy and operate. The other is the set of tools a gateway exposes to your agents once it is running: the actual callable functions, their access rules and their usage limits. A gateway with a bad tool catalogue is just a faster way to hand an agent too much power, so both halves deserve scrutiny.

For a concrete example of the second meaning, BusinessMCP publishes a generated reference of every tool its hosted MCP endpoint exposes. It lists 115 platform tools and 48 tools from connected services, marks which ones need approval before they act, and states plan gates and per-call prices where they exist. It is a useful model of what a transparent catalogue looks like, whether or not you ever use the product. The wider BusinessMCP managed MCP hosting page explains the hosted endpoint, its Bearer-key access control and the plans.

If you are new to the idea, read what an MCP gateway is first, then how it differs from an API gateway. If you are already choosing, the comparison of the best MCP gateway options and the evaluation checklist will save time. Teams with compliance obligations should start with MCP gateway security and audit logging.

What should you actually ask of any gateway? Four things cover most of the risk:

  • Who can call it? Per-caller keys or OAuth, not one shared secret.
  • What can each caller reach? A deny-by-default tool allowlist, not the whole catalogue.
  • What happens on risky actions? Sending email or spending money should wait for a human.
  • Can you reconstruct an incident? Every call logged with caller, tool and arguments.

A gateway is not a silver bullet. It cannot make a malicious server honest, and it cannot stop an agent from being talked into misusing tools it is allowed to call; BusinessMCP’s own guide to MCP prompt injection is blunt about that limit. What it can do is shrink the blast radius and give you a record.

We are an independent publication. Where a product is mentioned we say what it does, who it suits and what it does not do, and we link to primary pages so you can check. Features and plans change, so confirm details on the vendor’s page before you commit.

All articles

Fundamentals

Security

Choosing a gateway

Access control

Operations

Frequently asked questions

What are MCP gateway tools?

The phrase covers both the gateway software that proxies MCP traffic and the callable tools a gateway exposes to agents. Both need access control, logging and sensible limits.

Do I need an MCP gateway for one local server?

Usually not. A single local stdio server run by one person rarely needs a gateway. Gateways pay off when several people, machines or assistants share servers and credentials.

Is an MCP gateway the same as an API gateway?

No. An API gateway routes and limits HTTP calls. An MCP gateway also deals with tool discovery, per-tool permissions and the fact that a model decides which tools to call.

Are there open source MCP gateways?

Yes. Docker and Microsoft both publish open source gateway projects, and BusinessMCP publishes MIT-licensed hardening utilities for gateways. They differ a lot in scope, so read each README.

Can a gateway stop prompt injection?

Not by itself. It can limit which tools an injected instruction can reach and gate risky actions behind approval, but it cannot guarantee a model will never be manipulated.