Guide

MCP Gateway: What It Is, 7 Compared, and Who Needs One

12 min read·
MCP Gateway: What It Is, 7 Compared, and Who Needs One

Quick answer: An MCP gateway is one front door for all the MCP servers your AI agents use. Every tool call passes through it, so it can check who's asking, allow or block each tool, and log what happened. You need one when a team shares many servers and needs that control. One person with a few servers usually doesn't.

Questions this page answers

  • What is an MCP gateway, and what does it actually do?
  • How is it different from an MCP server, an API gateway, an AI gateway or a tool router?
  • Which MCP gateways exist, which are open source, and what do they cost?
  • Do you need one at all?
  • What should you use if you wanted more services, not more control?

Every vendor fact and price below was checked on the vendor's own page on November 13, 2026.

What is an MCP gateway?

An MCP gateway is a proxy that sits between AI clients (Claude Code, Cursor, an internal agent) and the MCP servers they call. Clients connect to the gateway once, and the gateway handles routing, sign-in and access rules for every server behind it.

If you're new to the protocol, start with what MCP is. The short version: an MCP server is a program that offers tools to an agent, like "search tickets" or "send email".

Without a gateway, every client is configured for every server. Ten developers and eight servers means eighty connections, each with its own credentials. With a gateway, it's ten connections to one door.

Without a gateway each agent connects to every MCP server itself; with one, every call passes through a single door that checks who is asking and logs it.

Docker's docs describe the flow in three steps:

  1. The agent asks for a tool.
  2. The gateway starts the right server in a container if it isn't running.
  3. The gateway injects the credentials, applies its restrictions, and forwards the call.

One thing a gateway does not do is give your agent new abilities. As Mastra's roundup puts it, a gateway "does not replace the underlying systems that provide an agent's capabilities." It controls access to servers you already have. For new abilities, connect your agent to Bento instead.

How is an MCP gateway different from an MCP server, an API gateway, an AI gateway or a tool router?

An MCP server provides tools. An MCP gateway governs access to many servers. The other three either route different traffic or sell you the tools themselves.

TermWhat it doesExample
MCP serverOffers tools to an agent over MCPA GitHub or Slack server
MCP proxyForwards one client to one server, often bridging transportsMany open-source bridges
MCP router / aggregatorPuts many servers behind one endpoint and one tool listMCP Router desktop app
MCP gatewayA router plus sign-in, access rules and audit logsDocker MCP Gateway, agentgateway
API gatewayRoutes and secures ordinary HTTP APIs; it doesn't understand toolsKong Gateway without its MCP plugin
AI (LLM) gatewayRoutes prompts to models, with keys, budgets and failoverOpenRouter, LiteLLM's model proxy
Tool routerRents you paid APIs from many vendors on one balanceMonid, treg

The lines blur in practice. agentgateway, LiteLLM, Kong and TrueFoundry each handle both model traffic and MCP traffic in one product. If you came here looking for "OpenRouter for tools", our guide to tool routers like Monid and treg covers that category.

What does an MCP gateway actually do?

It does five jobs, and most products do all five to different depths:

  1. One endpoint. Clients connect to one URL and see one combined tool list, often with each tool name prefixed by its server's name.
  2. Sign-in. Users or agents authenticate once, usually through your company's identity provider. The MCP spec's sign-in flow is OAuth (the MCP spec's authorization section), and most gateways build on it.
  3. Per-tool access rules. A support agent may list tickets but not delete them. Kong's plugin, for example, returns only the tools a caller's role allows, and answers anything else with HTTP 403.
  4. Logs and audit. Every call, allowed or denied, is recorded, so you can answer "which agent did that, and when".
  5. Running servers safely. Some gateways also host the servers. Docker's MCP Gateway docs describe running each one "in isolated Docker containers with restricted privileges, network access, and resource usage."

Which MCP gateways can you pick?

Seven are worth a look: four open-source gateways you run yourself (three with paid tiers on top), Kong's paid plugin, and two hosted services. Prices and licences below are as stated on each vendor's own page, checked November 13, 2026.

GatewayWhat it isOpen source or hostedPricing (vendor's own page)
Docker MCP GatewayRuns each MCP server in its own container, starts it on first call, injects credentialsOpen source; runs on your machine with Docker Desktop's MCP Toolkit or the docker mcp CLIFree. The governance version (Docker AI Governance) is invite-only through sales
agentgatewayOne proxy for model, MCP and agent-to-agent traffic, with RBAC, policies, audit logs and spend capsOpen source (Apache-2.0), a Linux Foundation project; self-hosted, standalone or on KubernetesFree. An enterprise offering is linked; no price shown
Kong AI MCP ProxyKong Gateway plugin: per-tool access rules, an audit log of every attempt, REST APIs turned into MCP toolsSelf-hosted Kong Gateway 3.12+, or Kong's Konnect cloudPlugin page: "AI License Required", part of AI Gateway Enterprise. Pricing page: Konnect Plus "from $25/month plus usage" lists MCP traffic. Confirm with Kong
LiteLLMMCP gateway inside LiteLLM's proxy: one fixed endpoint, access by key, team or organisationOpen source, self-hostedFree. Enterprise adds SSO beyond 5 users, audit logs and finer access control; priced on usage, through sales
IBM ContextForgeRegistry and proxy that federates MCP, agent-to-agent and REST/gRPC APIs behind one endpointOpen source (Apache 2.0); self-hosted from PyPI or Docker, scales on KubernetesFree. No paid tier on the repo
TrueFoundryHosted MCP gateway with Okta or Azure AD sign-in, per-server RBAC and auditHosted, or in your own VPC, on-prem or air-gappedDeveloper $0 (up to 3 users, 10K calls per user a month); Pro $25 per user a month (20K calls per user); Enterprise custom
Cloudflare MCP server portalsMany MCP servers on one endpoint inside Cloudflare One, with per-portal tool lists, request logging and "Code Mode"Hosted by Cloudflare; your servers can be public or privateCloudflare One: Free (up to 50 users), $7 per user a month pay-as-you-go, or contract. The docs don't say which plan includes portals

Two honest caveats. Kong's plugin page and pricing page don't tell the same story, so ask before you budget. And "open source" means the software is free to run, not that the people and servers to run it are.

Which one should you choose?

Choose by where your servers live and who already runs your infrastructure. The gateway that fits your existing stack beats the one with the longest feature list.

  • Pick Docker MCP Gateway if your developers already use Docker Desktop and you want each local server isolated in a container, with nothing new to host.
  • Pick agentgateway if a platform team runs Kubernetes and wants one open-source proxy for model, MCP and agent-to-agent traffic, with policy and audit.
  • Pick Kong if Kong already fronts your APIs and you want MCP rules in the same place.
  • Pick LiteLLM if you already route model calls through LiteLLM and want MCP tools behind the same keys and teams.
  • Pick IBM ContextForge if you want an open-source registry that also exposes your existing REST and gRPC services as tools.
  • Pick TrueFoundry if you want a hosted gateway with SSO and RBAC and a free tier to try it on.
  • Pick Cloudflare if your company already signs in through Cloudflare One and you want MCP access under the same rules.

Do you need an MCP gateway?

Probably not yet, unless several people or agents share several servers and someone has to answer for what they did. Even Mastra's roundup of gateways answers its own "Do I need one?" with "Not necessarily."

Count how many of these are true:

  1. More than one team or agent uses the same MCP servers.
  2. Some tools must be allowed for some callers and blocked for others.
  3. You need a record of every tool call for security or compliance.
  4. Credentials for each server are currently copied into each developer's config.

Two or more, and a gateway will earn its keep. One or none, and it's another service to run for a problem you don't have. Mastra's MCP gateway roundup covers the evaluation criteria in more depth if you're past this point.

When don't you need one?

Skip the gateway in three cases: you're one person with a handful of servers, your tool list is so long it's eating the agent's context, or what you actually want is more services rather than control over the ones you have.

One person, a few servers. Connect directly. In Claude Code that's one claude mcp add command per server; see our guide to adding an MCP server to Claude Code. A gateway adds a hop, a login and a thing to keep running.

Too many tools. Fifty tools in the context window is a cost problem, not a governance problem. Often the better fix is to let the agent write code instead of calling dozens of tools. Cloudflare's portals reach the same conclusion: their Code Mode collapses every tool into two, one to search and one to run code.

You wanted more, not control. People often hear "gateway" when what they're after is one connection that gives their agent search, data, media and their apps. A governance gateway won't add any of that. It only controls the servers you already run.

What if you want the services, not the gateway?

Then skip the governance layer and connect to an endpoint that already has the services behind it. That's what Bento is: the gateway to every service your agent needs.

Claude Code, Codex, Cursor or OpenClaw connect to Bento over MCP or the CLI. They then get AI models, web search and scraping, SEO data, company and people data, social data, voice and music, image, video and 3D generation, places to run code, and 3,000+ apps such as Gmail, Slack and Notion.

It's one subscription with no per-vendor keys to collect: Starter is $20 a month, Pro $50 and Max $200, and every new account gets $5 of free credits (checked November 13, 2026). See Bento's plans.

The limit, plainly: Bento doesn't sit in front of MCP servers you run. It adds no RBAC, policy or audit logs to them. If that's your job, pick a gateway from the table above. And if you're shipping your own agent product to customers, you still need a gateway or integration layer inside your own code.

Frequently Asked Questions

Is an MCP gateway the same as an MCP proxy?

Not quite. A proxy forwards one client to one server, often translating between transports. A gateway puts many servers behind one endpoint and adds sign-in, access rules and logs.

Is there a free, open-source MCP gateway?

Yes. Docker MCP Gateway, agentgateway (Apache-2.0), IBM ContextForge (Apache 2.0) and LiteLLM's gateway are all open source and free to run yourself. Their governance or enterprise tiers are sold separately, and you still pay to host them.

Does Claude Code need an MCP gateway?

No. Claude Code connects to MCP servers directly, one claude mcp add per server. A gateway helps only when a team needs shared credentials, per-tool rules or an audit trail across those servers.

Is an AI gateway the same as an MCP gateway?

No. An AI gateway routes prompts to language models and manages model keys and budgets. An MCP gateway routes tool calls to MCP servers. Several products, including agentgateway, LiteLLM, Kong and TrueFoundry, do both.

Can one MCP gateway connect to multiple MCP servers?

Yes, that's the point of one. Clients see a single endpoint with a combined tool list, and the gateway sends each call to the server that owns the tool.

Make anything. One subscription.

Posts, reels, ads, apps and games, made with the best AI.

Free credits to try it all. Nothing to set up.

Related guides