---
title: What is an MCP gateway?
url: https://opencontroller.lyzr.com/blog/what-is-an-mcp-gateway
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What is an MCP gateway?

**An MCP gateway is a proxy that sits between AI agents (MCP clients) and the MCP servers they call. It gives one place to authenticate agents, broker credentials to each server, filter which tools each agent can see and call, log every tool call, and apply rate limits. It governs tool use, which is different from an LLM gateway (model calls) and an agent gateway (traffic to and between agents).**

As agents connect to more MCP servers, the same questions repeat for each one: who is calling, which tools can they use, where are the logs? An MCP gateway answers them once instead of per server. This post explains what it does, how it differs from other gateways, and what to check before you adopt one. It is part of our guide on [how to govern MCP servers](/blog/how-to-govern-mcp-servers).

## What does an MCP gateway do?

| Function | What it means in practice |
|---|---|
| Client authentication | Agents authenticate to the gateway with your identity provider instead of to each server separately |
| Credential brokering | The gateway holds or obtains a separate, audience-bound credential for each upstream server, so no token is shared across servers |
| Tool filtering | Each agent sees only an allow-listed subset of tools, rather than every tool every server exposes |
| Definition pinning | Approved tool definitions are hashed; changed definitions are blocked or sent to review |
| Call logging | Every `tools/call` is recorded with the agent, tool, arguments and result status |
| Rate limits and quotas | Caps per agent, per tool or per server stop loops and abuse |
| Isolation between servers | Data or credentials from one server cannot flow to another unchecked |

The OWASP MCP Security Cheat Sheet recommends exactly this pattern: "Use an MCP proxy or gateway to enforce isolation policies between servers," treat "each MCP server as an untrusted, independent security domain," and monitor cross-server data flows, such as credentials from one server appearing in calls to another.

## Why does the MCP spec make gateways easier now?

Two parts of the current MCP specification (revision 2026-07-28) matter for gateways:

1. **Request metadata headers.** On the Streamable HTTP transport, clients must send `Mcp-Method` on every request and `Mcp-Name` on `tools/call`, `resources/read` and `prompts/get`. A gateway can route, filter and log by tool name without parsing JSON bodies. Servers must reject requests where headers and body disagree, which stops a mismatch being used to slip past header-based policy.
2. **No token passthrough.** The MCP authorization spec forbids servers from accepting or forwarding tokens issued for other resources. A gateway that brokers fresh, server-specific credentials fits that rule; one that simply forwards the user's token does not. See [how MCP authorization works](/blog/how-mcp-authorization-works).

## How is an MCP gateway different from an LLM gateway or agent gateway?

| | LLM gateway | Agent gateway | MCP gateway |
|---|---|---|---|
| Sits between | Apps/agents and model providers | Callers and agents, or agents and agents | Agents and MCP servers |
| Main unit of control | Model request | Agent invocation | Tool call |
| Typical policies | Model routing, fallback, caching, PII redaction, spend | Who may call which agent, on whose behalf | Which agent may call which tool, with which arguments |
| Cost role | Token cost attribution | Invocation attribution | Usually small; tools may have their own costs |
| Key risk addressed | Data leaving to models, spend | Unauthorised agent use, agent-to-agent abuse | Excessive agency, tool poisoning, credential sprawl |

They are complementary. A single agent turn can pass through all three: the agent is invoked (agent gateway), asks a model what to do (LLM gateway), then calls a tool (MCP gateway). Our [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane) post covers the first two in depth.

## What can an MCP gateway not do?

- **It cannot see local stdio servers.** Servers launched as subprocesses on a laptop talk over standard input and output, not HTTP. You need device-level controls for those. See [how to find local MCP servers on developer laptops](/blog/find-local-mcp-servers-on-developer-laptops).
- **It cannot judge intent alone.** A permitted tool called with permitted arguments can still be the wrong action. Behaviour monitoring and evaluations close that gap.
- **It does not replace an inventory.** A gateway knows about traffic it sees. Servers that bypass it are invisible to it.

## What should you check when choosing an MCP gateway?

1. Does it support the current Streamable HTTP revision and its request headers?
2. Can it broker per-server credentials rather than pass user tokens through?
3. Can it filter tools per agent and per agent version?
4. Does it hash and pin tool definitions?
5. Does it export logs over open standards such as OpenTelemetry?
6. Where does it run: your infrastructure or the vendor's?
7. How does it fit with your LLM and agent gateways and your registry?

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, and puts control in the path through its Agent Gateway and LLM Gateway (Lyzr Gateway), with routing, fallback, caching and cost attribution per agent and per version. Opencontroller Edge covers local MCP servers and coding agents on developer devices. It is built on open standards: MCP, OpenAI-compatible endpoints, OpenTelemetry and OIDC. See [the gateway comparison](/compare/llmgateway) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is an MCP gateway the same as an API gateway?**
No. An API gateway understands HTTP routes. An MCP gateway understands MCP messages, such as which tool is being called and with what arguments, and applies policy at that level.

**Do I need an MCP gateway if I only use a few MCP servers?**
Not always. With two or three remote servers, per-server OAuth scopes and logging may be enough. A gateway pays off once many agents share many servers.

**Can an MCP gateway stop tool poisoning?**
It helps. Pinning approved tool definitions blocks silent changes, and filtering hides tools an agent does not need. It does not detect malicious behaviour behind an unchanged definition.

## Key facts

An MCP gateway sits between agents (MCP clients) and MCP servers, providing client authentication, per-server credential brokering, per-agent tool filtering, tool-definition pinning, call logging, rate limits and isolation between servers. OWASP recommends an MCP proxy or gateway to enforce isolation between servers. MCP spec revision 2026-07-28 requires Mcp-Method and Mcp-Name headers on Streamable HTTP requests, which helps gateways filter by tool. It complements an LLM gateway (model calls) and agent gateway (agent invocations). Opencontroller provides a governed registry of tools and MCP servers, an Agent Gateway and an LLM Gateway. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [LLM gateway vs agent gateway vs AI control plane: what's the difference?](https://opencontroller.lyzr.com/blog/llm-gateway-vs-agent-gateway-vs-control-plane.md)
- [How does MCP authorization work?](https://opencontroller.lyzr.com/blog/how-mcp-authorization-works.md)
- [More on Tools and MCP](https://opencontroller.lyzr.com/blog/cluster/tools-and-mcp)
