How to Deploy an AI Agent: A Step-by-Step Guide
Learn how to deploy an AI agent: pick an execution pattern, choose a host, and launch safely, with real examples from LangGraph, AWS, Google Cloud and Salesforce.
7 min read
An MCP gateway sits between AI agents and MCP servers. Here are 8 options compared, from open source projects like Docker and Microsoft's to managed platforms like Composio and Kong.

Run more than a couple of MCP servers and a new problem shows up fast: every agent needs its own credentials, its own list of endpoints, and its own way of finding out which tools exist. An MCP gateway is the piece of infrastructure built to make that problem go away, and in 2026 there are enough of them, open source and commercial, that picking one is its own research project.
An MCP gateway is a server that sits between AI agents and one or more Model Context Protocol (MCP) servers, exposing them behind a single endpoint. It handles authentication, routes each tool call to the right backend server, and logs what happened, so an agent connects once instead of managing a separate connection, credential, and access policy for every tool it uses.
The gap this post fills: every page ranking for "mcp gateway" right now is a single vendor explaining its own product, or Microsoft's and IBM's bare GitHub repositories with no comparison attached. None of them puts the open source projects next to the managed platforms so a developer can see what each one actually does before picking.
Kong's explainer, published 10 June 2026, frames the problem as the "too many endpoints" problem: without a gateway, every client maintains its own connection to every MCP server, so credentials multiply and a single policy change has to be pushed to every client by hand. The gateway collapses that into a hub-and-spoke shape: clients know one URL, and the gateway handles routing, authentication, and policy from there.
Composio's own comparison guide, "10 Best MCP Gateways for Developers in 2026" (28 July 2026), calls this the "N×M problem": N agents each needing to talk to M tools means N×M integrations to secure and maintain without a shared layer in between. A gateway turns that into N+M: every agent connects to the gateway once, and every tool is registered with the gateway once.
Kong's piece adds a concrete trigger point for when to bother: more than two or three MCP servers, a compliance requirement that demands centralized logging, more than one team needing isolated access, or a move from a prototype into production. Below that line, a direct connection is simpler.
| Gateway | What it is | Best for | Price |
|---|---|---|---|
| Docker MCP Gateway | Open source proxy bundled into Docker Desktop and the Docker Agentic Platform | Teams already standardized on Docker tooling | Free, open source |
| Microsoft MCP Gateway | Open source (MIT), Kubernetes-native reverse proxy and control plane with a management portal | Platform teams running MCP servers on Kubernetes or Azure | Free, open source |
| IBM ContextForge (mcp-context-forge) | Open source (Apache 2.0) gateway, registry, and proxy federating MCP, A2A, and REST/gRPC behind one endpoint | Teams unifying MCP with existing A2A or REST infrastructure | Free, open source |
| AWS MCP Gateway & Registry | Open source AWS Labs project pairing a gateway with a searchable server registry and access control | AWS-centric teams that want a catalog plus governance | Free, open source (AWS infra costs apply) |
| Kong AI Gateway (MCP Gateway) | Enterprise API gateway vendor's MCP layer, built into Kong AI Gateway | Enterprises already running Kong for API management | Paid, enterprise pricing |
| TrueFoundry MCP Gateway | Managed SaaS with a free tier and VPC, on-prem, or air-gapped deployment options | Teams wanting a managed registry with a free starting tier | Freemium (30,000 tool calls a month free) |
| Composio MCP Gateway | Managed gateway bundled with 1,000+ pre-built tool integrations and SOC 2 Type II / ISO 27001 certification | Teams that want integrations and governance from one vendor | Paid, custom pricing |
| Boomi MCP Gateway (formerly Lunar.dev MCPX) | Part of Boomi's AI Gateway product line, paired with an LLM Gateway for combined model- and tool-traffic governance | Teams already on Boomi's integration platform | Paid, enterprise pricing |
Docker's MCP Gateway is described in its own docs as Docker's open source solution for orchestrating MCP servers, acting as a centralized proxy between clients and the servers it fronts. It ships as part of the Docker MCP Catalog and Toolkit, so a team that already builds and runs containers with Docker gets a gateway without adding a new vendor to the stack. It is Compose-first: servers are defined and launched the way any other Docker service would be.
Microsoft's mcp-gateway is MIT-licensed and built specifically for Kubernetes, splitting itself into a data plane that routes each MCP request to a ready server instance and a control plane with RESTful APIs for deploying, updating, and deleting servers. It ships with a React-based management portal for listing, editing, and testing adapters and tools in a browser, and supports one-click deployment to Azure. As of this writing, the project requires MCP spec version 2026-07-28 clients and adapters; its own migration guide describes this as a breaking change with no legacy session or protocol-downgrade support.
IBM's mcp-context-forge describes itself as an AI gateway, registry, and proxy that sits in front of any MCP, A2A, or REST/gRPC API, exposing one unified endpoint with centralized discovery and guardrails. It is Apache 2.0 licensed and, as of 15 October 2026, has reached 4.6k GitHub stars and 925 forks. Because it federates protocols rather than just MCP, it fits teams that already run A2A agents or REST services alongside MCP servers and want one gateway in front of all three instead of three separate layers.
AWS's open source blog post "Governing AI Assets at Scale with MCP Gateway and Registry", published 17 June 2026 by Amit Arora and Omri Shiv, describes the agentic-community/mcp-gateway-registry project as pairing a gateway with a React-based UI for fine-grained access control alongside an AI Registry Tools MCP server. The project integrates Cisco AI Defense's open source mcp-scanner, a2a-scanner, and skill-scanner, and connects to GoDaddy's Agent Name Service (ANS) for naming agents across the registry. It is the option to reach for when a team is already standardized on AWS and wants a catalog of servers, not just a routing layer.
Kong extends its existing API gateway into MCP rather than building a separate product. Its own explainer lists OAuth 2.1, OpenID Connect, and SAML support, tool-level access control lists that let a security-scanner agent see two tools out of forty a backend server exposes, and a feature that converts existing REST API endpoints into MCP-compatible tools without manual code. Kong's pitch for teams already running it for API management is that MCP, LLM, and A2A traffic then sit on one control plane instead of three.
TrueFoundry's MCP Gateway is a managed product offering unified access to MCP servers with a registry attached, and it is one of the few gateways on this list with a published free tier: 30,000 tool calls a month, no credit card required. For teams past the free tier, TrueFoundry advertises VPC, on-prem, and air-gapped deployment, aimed at the enterprises its site names as customers, including Cargill and Mavenir. Composio's comparison guide separately credits TrueFoundry with single-digit-millisecond routing overhead, among the lowest latency figures it found publicly disclosed.
Composio positions its MCP Gateway as a managed control plane bundled with its existing catalog of more than 1,000 tool integrations, meaning a team gets a gateway and a large set of pre-built connectors from the same vendor rather than wiring its own MCP servers first. Composio's own comparison piece claims SOC 2 Type II and ISO 27001 certification with configurable log retention and CSV export, and SCIM-driven team provisioning. The same guide is explicit that it is written by Composio and recommends Composio as the default for most teams, so its claims about itself are worth weighing against the open source and competing managed options above.
Lunar.dev's MCPX product is now documented under Boomi's help site as the MCP Gateway component of Boomi's broader AI Gateway, alongside a companion LLM Gateway for model traffic and a Skills feature that packages a SKILL.md-style capability with its own tools and prompts. The page, last updated 31 August 2026, describes the MCP Gateway as registering multiple MCP servers under one governed endpoint, enforcing per-user scoped permissions, and sharing unified policy and identity with the LLM Gateway. Composio's comparison guide separately lists Lunar.dev MCPX among the lowest-latency gateways it benchmarked.
Composio's comparison guide lays out nine decision criteria for a production gateway: security posture (OAuth 2.1 with PKCE, PII redaction), deployment model (managed SaaS versus self-hosted versus VPC), compliance and audit trail, identity and SCIM provisioning, MCP-specific threat handling (tool poisoning, rug-pull updates, cross-server shadowing), p95 latency, developer experience, protocol and transport interoperability, and pricing model.
That last one deserves its own sentence: agents are chatty. A single user request can fan out into dozens of tool calls, so per-task or per-call pricing that looks reasonable for human-triggered automation can turn unpredictable once an agent is driving it. Model your expected call volume against each vendor's pricing before shortlisting, not after.
For data residency or regulated-industry constraints, rule out pure SaaS gateways first and only then compare the self-hosted and VPC options against each other. For teams that only need breadth across everyday business apps rather than MCP-specific governance, Composio's guide also points to Zapier's MCP offering, which connects to more than 8,000 apps, as an option worth weighing against the MCP-native gateways on this list.
An MCP gateway is a server that sits between AI agents and one or more MCP servers, exposing them through a single endpoint. It handles authentication, routes each tool call to the correct backend server, and logs activity, replacing a setup where every agent would otherwise connect to, and hold credentials for, every MCP server individually.
Not necessarily. Kong's own guidance suggests starting with direct client-to-server connections and adding a gateway once complexity shows up: more than two or three servers, a compliance requirement that demands centralized logging, multiple teams needing isolated access, or a move from a prototype into production.
An MCP server implements the actual tools and business logic an agent calls, such as a database query or a CRM lookup. An MCP gateway never provides tools itself; it manages access to the servers that do, handling authentication, routing, and policy enforcement in front of them.
Yes. Docker MCP Gateway, Microsoft's mcp-gateway (MIT licensed), IBM's mcp-context-forge (Apache 2.0), and AWS Labs' mcp-gateway-registry project are all open source with no license cost, though self-hosting still carries infrastructure and operational overhead.
A gateway routes live traffic between agents and MCP servers at request time. A registry is a catalog of which MCP servers exist and what tools they expose, used for discovery rather than routing. Several projects on this list, including AWS's mcp-gateway-registry and TrueFoundry's product, bundle both into one offering.
Browsing the MCP servers listed on AI Agents Listing is a reasonable next step once a gateway choice is narrowed down, since the servers a team picks will shape which gateway features actually matter.
Source: Kong, "What is a MCP Gateway?"
More options for teams building out their developer tooling stack are on AI Agents Listing's developer tools category.
One email a week. New agents, MCP servers and skills, and what is actually getting traction.
Learn how to deploy an AI agent: pick an execution pattern, choose a host, and launch safely, with real examples from LangGraph, AWS, Google Cloud and Salesforce.
7 min read
AI agent pricing models explained with real numbers from Salesforce Agentforce, Intercom's Fin AI Agent, and Anthropic's Claude API pricing pages.
6 min read
MCP vs Agent Skills, compared on how each connects an AI agent to the outside world: MCP as a client-server protocol for live tools, Agent Skills as local Markdown instructions.
7 min read