/aienm.

Guide

MCP Gateway: What It Is, What It Should Enforce, and How to Choose One

One endpoint between your agents and every MCP server they use: where identity, tool permissions and audit logs get enforced, and what to check before you pick one.

By AIENM Editors · Reviewed

What an MCP gateway is

An MCP gateway is a single endpoint that sits between MCP clients (agents, IDE coding assistants, chat apps) and the MCP servers that expose tools, prompts and resources to them. Instead of configuring every client with every server, each client connects once to the gateway. The gateway decides which servers and tools that caller may see, routes each request to the right server, attaches the credentials the server needs, and records what happened.

It plays the role an API gateway plays for REST traffic, with one difference that matters for governance: it speaks the Model Context Protocol. It can read which tool is being called and with which arguments, so policy can be written per tool rather than per URL. To the client, the gateway looks like one MCP server; to each upstream server, it looks like one MCP client.

If you are still deciding what an MCP server contributes to an agent in the first place, start with What an MCP Server Does in an Agentic Stack, then come back here.

Why enterprises put a gateway in front of MCP servers

The protocol standardises how a client calls a tool. It deliberately leaves most of what a security or platform team needs to the deployment:

  • Authorization is optional. The MCP specification says so in as many words, and where servers do implement it, it defines an OAuth flow, not a role model. Who may call which tool is left to you. Role-Based Access Control in Enterprise MCP Server Deployments at Scale covers that gap in depth.
  • Configuration multiplies. Without a gateway, every client holds its own list of servers and its own credentials for each. Revoking access or retiring a server means finding every copy.
  • There is no shared inventory. Teams stand up servers independently, and nobody can answer which servers are running, who owns them and which version an agent is calling.
  • Audit is scattered. Each server logs (or doesn't) in its own format, so tracing a write back to the person whose agent made it is a forensic exercise.
  • Tool lists crowd the context window. Every connected server adds its tool definitions to the model's context. Past a point, more tools means worse tool selection and higher cost, which is why several gateways filter or search tools rather than listing all of them.

What a gateway should enforce

Gateways are sold on very different feature lists. These are the controls that decide whether one actually governs agent traffic, roughly in the order they matter:

  1. Caller identity on every request. The gateway should authenticate each request against your identity provider and know which person (or which workload) is behind it, not only which agent. Non-Human Identity Management for Enterprise AI Agents explains why the agent's own identity and the user's both need to travel.
  2. No token passthrough. The MCP authorization spec requires servers to accept only tokens issued for them as the audience, and says servers must not accept or transit any other tokens. A gateway that forwards the client's token to every upstream server breaks that rule. Look for one that obtains its own credential for each upstream server (from a vault, a credential provider or a token exchange).
  3. Permissions per tool, not per server. A server that offers both read and delete operations needs the delete tool restricted separately. Check that policies can name individual tools and, ideally, constrain their arguments. See Tool Calling and Action Permissions in AI Agents.
  4. A catalog of approved servers. Each server registered with an owner, a version and an approval record, so the gateway only routes to what has been reviewed. Governed AI Registries for Shared Agent Skills covers what that registry should hold.
  5. An audit record per tool call. Who called, through which agent, which tool, with which arguments, what came back and whether policy allowed it, exported to the SIEM you already run.
  6. Isolation for the servers it runs. Where the gateway also hosts servers, each should run with restricted filesystem, network and resource access, so one compromised server cannot reach the others.
  7. Tool filtering. Expose to each caller only the tools their role and task need, both to shrink the attack surface and to keep tool lists out of the context window.
  8. Rate limits and spend controls per user, team or key, so a looping agent cannot exhaust a downstream system.
  9. Content inspection. Tool results are untrusted input to the model. Some gateways can redact secrets or flag injection patterns in what servers return; see Prompt Injection Risks in Production Enterprise Agents for what that does and does not catch.

The main MCP gateway options

The options below are the ones people search for most by name. Each summary is drawn from the vendor's or project's own documentation as of November 2026; features and packaging change quickly, so confirm against the linked docs before you shortlist. Pricing is not listed: check each vendor's pricing page or ask their sales team.

MCP gateway options, as described in their own documentation
OptionHow it runsWhat its docs say it doesCheck
Docker MCP GatewayOpen source; runs automatically with Docker Desktop's MCP Toolkit, or as a CLI plugin on Docker EngineRuns each MCP server in an isolated container with restricted privileges, network access and resources; starts servers on demand, injects their credentials, and logs and traces calls. Profiles decide which servers a client sees. The version inside Docker AI Governance is invite-only.Docker docs
Microsoft MCP GatewayOpen source (MIT), self-hosted on Kubernetes, with templates for AzureA reverse proxy and control plane: deploys, updates and removes MCP servers, registers individual tools behind a tool router, and authorizes with Entra ID app roles. The current version requires clients on the MCP 2026-07-28 revision.GitHub
Amazon Bedrock AgentCore GatewayManaged AWS serviceCombines Lambda functions, OpenAPI specs, Smithy models and remote MCP servers into one virtual MCP server, with semantic tool search. Inbound auth by OAuth (JWT) or IAM; outbound credentials held in AgentCore credential providers.AWS docs
Kong AI MCP ProxyA plugin for Kong Gateway, part of Kong's AI Gateway Enterprise offeringBridges MCP and HTTP, so MCP clients can call existing Kong-managed APIs as tools or reach upstream MCP servers through Kong.Kong docs
LiteLLM MCP GatewayPart of the LiteLLM Proxy, self-hostedOne /mcp endpoint for every registered MCP server, with access controlled by API key and team, sharing the same auth, rate limits and usage tracking the proxy applies to model calls.LiteLLM docs
IBM ContextForgeOpen source (Apache 2.0), self-hostedA gateway, registry and proxy in front of MCP, A2A and REST/gRPC APIs, with one endpoint, centralised discovery, guardrails and plugins.GitHub
MCP Gateway & Registry (agentic-community)Open source (Apache 2.0), self-hostedA gateway and registry with OAuth sign-in through Keycloak or Entra ID and dynamic tool discovery, aimed at both autonomous agents and coding assistants.GitHub

Several broader agent platforms and integration vendors also ship a gateway as one part of a larger product. Enterprise MCP Server Providers With Built-In Governance Controls and Agent Registry and Tool Catalog Solutions for Enterprise AI Governance compare those.

Four ways teams deploy one

On the developer's machine

A local gateway (Docker's is the common example) gives each developer one connection for their IDE assistant and runs servers in containers instead of as unvetted local processes. It fixes sprawl and isolation on laptops; it does not give the organisation a central audit trail on its own.

Self-hosted, as shared infrastructure

A platform team runs the gateway on its own Kubernetes cluster (Microsoft's, ContextForge or the agentic-community project) and every team's agents point at it. You keep the data and logs in your network and own the upgrades.

Managed, inside your cloud

AgentCore Gateway and similar services take the operations work away and tie into the cloud's own identity and logging. The trade is that policy and logs live in that provider's model.

On the API gateway or LLM proxy you already run

If Kong already fronts your APIs or LiteLLM already fronts your model traffic, adding MCP there keeps one place for auth, rate limits and usage. Check that its MCP support covers per-tool policy and not only routing. MCP vs API for Enterprise Agent Integrations covers when an existing API is the better integration path anyway.

Questions to ask in a pilot

  • Does the user's identity reach the tool call, or does every call run as the agent's service account?
  • Can a policy allow one tool on a server and deny another? Can it restrict arguments?
  • Does the gateway ever forward a client's token upstream? How are upstream credentials stored and rotated?
  • Which MCP protocol revisions does it support? The 2026-07-28 revision changed how requests are routed, and some gateways support only newer or only older clients.
  • What exactly is logged for each call, how long is it kept, and can it be streamed to your SIEM?
  • Can it register servers with an owner, version and approval status, and block anything unregistered?
  • Can it turn existing REST APIs into tools, or does every integration need its own MCP server?
  • Where does it run and where do logs and credentials live, given your data residency rules?
  • What happens to agents when the gateway is down: fail closed, or fall back to direct connections?

For hardening the servers behind the gateway, see How to Secure an MCP Server in a Production Environment and Scoped MCP Design for Enterprise Teams.

Frequently asked questions

Is an MCP gateway the same as an API gateway?

No, though some API gateways now add MCP support. An API gateway sees HTTP requests to URLs; an MCP gateway understands MCP messages, so it can list, filter and authorize individual tools and record the arguments of each call.

What is the difference between an MCP gateway and an MCP registry?

A registry is the inventory: which servers exist, who owns them, which version is approved. A gateway is the enforcement point that traffic passes through. Several products combine both, and a gateway is most useful when it routes only to servers the registry has approved.

Is there an open-source MCP gateway?

Yes. Docker's MCP Gateway, Microsoft's MCP Gateway (MIT), IBM ContextForge (Apache 2.0) and the agentic-community MCP Gateway & Registry (Apache 2.0) are all published as open source, and LiteLLM's proxy includes one. Check each project's licence and which features sit in a paid edition.

Do I need a gateway if I only run a few MCP servers?

Not necessarily. With a handful of servers, one team and read-only tools, configuring clients directly can be enough. A gateway starts paying off once several teams share servers, tools can write to business systems, or someone has to answer who called what.

Does the MCP specification define a gateway?

The specification defines clients, servers and an authorization flow; a gateway is a deployment pattern built from those roles. Its rules still apply to the gateway: to clients it is an MCP server, so it must validate tokens issued for it, and it must not pass those tokens on to the servers behind it.

Further reading on AIENM

Sources

  1. Model Context Protocol specification (2026-07-28): Authorization
  2. Docker: MCP Gateway
  3. microsoft/mcp-gateway on GitHub
  4. Amazon Bedrock AgentCore Gateway: core concepts
  5. Kong: AI MCP Proxy plugin
  6. LiteLLM: MCP Overview
  7. IBM/mcp-context-forge on GitHub
  8. agentic-community/mcp-gateway-registry on GitHub