Skip to main content
Agentic AI Security 101
7 min read
October 24, 2026

What Are the Best MCP Gateways in 2026?

By SecureAuth Technology Team

Summarize with
Table of Contents

Loading sections...

Key takeaways

  • Enterprise MCP gateways in 2026 fall into four groups: API and AI gateways that added MCP, identity-first gateways, open source self-hosted gateways and managed MCP platforms.
  • Compare them on five points: an identity per agent run tied to the user, policy on each tool call, credentials, audit and deployment.
  • Open source options include Docker MCP Gateway, IBM ContextForge, Microsoft MCP Gateway, Obot, agentgateway and Agent Router (formerly Envoy AI Gateway).
  • Three trial tests show whether a gateway decides each call: revoke one agent mid-task, deny one tool for one user, and allow a server's safe tools while denying the destructive ones.
  • Where a gateway only routes and signs users in, add an authorization layer that checks every tool call against the agent and the delegating user.

Short answer. The leading enterprise MCP gateways in 2026 fall into four groups: identity-first gateways (SecureAuth Agent Authority, Okta Agent Gateway, Aembit, Pomerium, Runlayer, Cloudflare MCP server portals), API and AI gateways that added MCP (Kong AI Gateway, Azure API Management, agentgateway), open source gateways you host yourself (Docker MCP Gateway, IBM ContextForge, Microsoft MCP Gateway, Obot) and managed MCP platforms (TrueFoundry, Portkey, MintMCP). Most handle routing, sign-in and logging. Before you choose, check whether the gateway gives each agent run its own identity tied to the user and checks every tool call, down to its arguments.

The MCP gateways enterprises compare in 2026

An MCP gateway gives every agent one controlled path to your MCP servers. Each entry below summarizes what the vendor documents today and who it tends to fit, with a link to the source. A few products also cover the path from agents to AI model providers, such as Agent Router and SecureAuth Agent Authority.

Identity-first gateways

Identity-first gateways start from who is calling. They authenticate the person, and in some cases the agent, through your identity provider before any tool call, then use that identity for tool policy, credentials and audit.

  • SecureAuth Agent Authority (MCP Authority and LLM Authority). Agent Authority works downstream of the identity provider you already have and covers both paths an agent uses: MCP Authority for tool calls and LLM Authority for requests to AI model providers. On the MCP path, each agent run gets its own identity tied to the delegating user, each tool call is checked against policy down to its arguments, short-lived OAuth tokens replace API keys, a denied tool is left out of that user's tool list, and denied calls are recorded. On the model path, LLM Authority allows only approved models per team, caps spend per team, strips sensitive data from prompts and responses, and logs every call with the agent and the person behind it. Fits: Security teams that want one access-control layer for agents' tool calls and model calls, without replacing their IdP. See MCP Authority and LLM Authority.
  • Okta Agent Gateway. Okta describes Agent Gateway as an identity-native proxy that combines tools from several remote MCP servers into one Okta-secured endpoint, enforces identity and policy on every tool call, and injects downstream credentials so the agent never sees them. Fits: Okta customers that want agent access to MCP tools managed in Okta. Source: Okta Agent Gateway.
  • Aembit MCP Identity Gateway. A data-plane gateway between AI agents and MCP servers. Users sign in through an IdP such as Okta, Entra ID or Google, and the token the agent presents identifies both the agent and the user. Policies authorize access per MCP server, and the gateway fetches server credentials per request so the agent never receives them. Fits: Teams that want the agent and the user checked at the gateway, with a managed or self-hosted option. Source: Aembit MCP Identity Gateway docs.
  • Pomerium. Pomerium sits between MCP clients and MCP servers, authenticates users through your IdP with OAuth 2.1, and lets policies match tool names alongside user, email or group. It injects upstream OAuth tokens so clients never see them, and logs every tool call with its parameters. Fits: Teams that want MCP servers behind IdP sign-in, with policies that name specific tools. Source: Pomerium MCP support.
  • Runlayer. Runlayer requires a credential on every MCP request, such as an SSO user token, OAuth token, agent token or API key, and applies RBAC, PBAC or ABAC rules to which users, groups or agents can reach each server and tool. It acts as an OAuth proxy for third-party services and logs every tool call with user, parameters and response. Fits: Companies that want a central MCP control point with SSO, tool-level rules and SIEM export. Source: Runlayer reverse proxy docs, Runlayer MCP security best practices.
  • Cloudflare MCP server portals. Portals put many remote MCP servers behind one endpoint that users reach through Cloudflare Access. Admins can turn off tools so they disappear from a portal's tool list, and Access logs the individual tool requests. Fits: Organizations already on Cloudflare One that want MCP access behind their zero-trust policies. Source: Cloudflare MCP server portals.

API and AI gateways that added MCP

  • Kong AI Gateway. Kong's AI MCP Proxy plugin applies allow and deny rules per tool to Consumers and Consumer Groups and can log audits. The AI MCP OAuth2 plugin, a Tech Preview that needs an AI Gateway Enterprise license, implements the MCP OAuth 2.0 authorization spec. Fits: Teams that already run Kong and want MCP under the same gateway, in Konnect or on-prem. Source: Kong MCP tool ACLs, Kong AI MCP OAuth2 plugin.
  • Azure API Management. API Management can expose any REST API it manages as an MCP server, or put an existing remote MCP server behind it, with JWT validation, rate limits, IP filtering and Azure Monitor logging. Policies currently apply to all tools in an MCP server rather than to individual tools. Fits: Azure shops that want MCP governed like their other APIs, including through the self-hosted gateway. Source: Azure API Management MCP overview.
  • agentgateway. An open source gateway, a Series of LF Projects that has joined the Agentic AI Foundation. Its MCP authorization uses CEL rules that decide which tools, prompts and resources a client can reach, and the rules can match JWT claims. Fits: Platform teams that want an open source, policy-as-code gateway, standalone or on Kubernetes. Source: agentgateway MCP authorization, agentgateway authorization.
  • Agent Router (formerly Envoy AI Gateway). The Envoy AI Gateway project, renamed Agent Router in September 2026, remains open source under Apache 2.0 and offers MCP endpoints alongside model routing, with Envoy as the data plane. Fits: Teams already on Envoy that want model and MCP traffic in one open source router, from a laptop to Kubernetes. Source: Agent Router joins the Agentic AI Foundation.

Open source, self-hosted gateways

  • Docker MCP Gateway. Docker's open source gateway runs each MCP server in an isolated container, injects the credentials it needs and includes built-in logging and call tracing. Profiles decide which servers clients can use. Fits: Developers and teams that want containerized MCP servers on Docker Desktop or Docker Engine. Source: Docker MCP Gateway docs.
  • IBM ContextForge. An Apache 2.0 registry and proxy that federates MCP, A2A and REST or gRPC services into one endpoint, with JWT and SSO sign-in, user-scoped OAuth tokens, plugins and OpenTelemetry tracing. Fits: Teams that want a self-hosted gateway they can extend, from a pip install to a Helm chart. Source: IBM ContextForge on GitHub.
  • Microsoft MCP Gateway. An open source reverse proxy and management layer for MCP servers on Kubernetes, with Entra ID authentication and basic application-role authorization for MCP servers and tools. Fits: Teams running MCP servers on AKS that want Entra sign-in in front of them. Source: Microsoft MCP Gateway docs, Microsoft MCP Gateway deployment.
  • Obot. An MIT-licensed MCP gateway with policies on which users, groups and agents can call which tools on which servers. OAuth tokens stay in the gateway, and every tool call is logged with its user, agent, server, arguments and outcome. Fits: Teams that want open source they can run on Kubernetes or Docker, with a managed option. Source: Obot MCP Gateway.

Managed MCP platforms

Managed MCP platforms give teams a curated catalog of approved MCP servers with credentials and access rules set centrally, run by the vendor or deployed in your own cloud, and often sit next to the vendor's AI model gateway. SecureAuth Agent Authority, listed under identity-first gateways, can fill this role too: MCP Authority offers a self-service catalog of vetted servers behind one governed connection, and LLM Authority covers the model path.

  • TrueFoundry MCP Gateway. Admins decide which tools within an MCP server each user can invoke. The gateway handles per-user OAuth2 with token refresh, client credentials or static keys, and its audit logs track which user made each request. Fits: Enterprises that want the gateway in their own VPC, on-prem or air-gapped, next to TrueFoundry's AI gateway. Source: TrueFoundry MCP gateway auth and security, TrueFoundry MCP Gateway.
  • Portkey MCP Gateway (now Prisma AIRS AI Gateway). Portkey, now part of Palo Alto Networks' Prisma AIRS AI Gateway, injects OAuth tokens, API keys or identity headers for each server, can enable or disable specific tools per user, and logs every tool call with the user, parameters and response. Fits: Teams that already route model traffic through Portkey and want MCP in the same place. Source: Portkey MCP Gateway docs.
  • MintMCP (now Mintive). A managed catalog of approved MCP servers with credentials configured centrally, role-based tool sets, the option to turn off destructive tools, and audit logs that trace calls to a person, agent, server and model. Fits: Companies that want a hosted catalog of ready-to-use MCP servers behind SSO. Source: MintMCP docs.

MCP gateway comparison: identity, tool policy, credentials, audit and deployment

Each cell reflects the vendor's public documentation as of October 24, 2026. "Not documented" means we didn't find it in those pages, not that the product can't do it, so ask the vendor during a trial.

GatewayIdentity per agent run, tied to the userPolicy on each tool callCredentialsAuditDeployment
SecureAuth MCP AuthorityYes, per agent run, tied to the delegating userEvery call, down to its argumentsShort-lived OAuth tokens instead of API keysDenied calls are recordedDownstream of your existing IdP
Kong AI GatewayNot documentedPer-tool ACLs by Consumer and Consumer Group; arguments not documentedMCP OAuth 2.0 plugin (Tech Preview, Enterprise) or API keysAudit logging optionKonnect or self-hosted
Azure API ManagementNot documentedPer server; policies apply to all tools in itJWT validation inbound; keys or OAuth to backendsAzure Monitor, Application InsightsAzure-managed; self-hosted gateway
agentgatewayNot documentedCEL rules per tool, prompt or resource, using JWT claims; arguments not documentedValidates JWTs from your IdPNot documentedOpen source; standalone or Kubernetes
Agent Router (formerly Envoy AI Gateway)Not documentedNot documentedNot documentedNot documentedOpen source (Apache 2.0); laptop to Kubernetes
Cloudflare MCP server portalsNot documentedTools turned on or off per portal; servers shown per user by Access policyOAuth, bearer tokens, per-user or admin credentialsAccess logs each tool requestCloudflare-hosted
Okta Agent GatewayNot documentedIdentity and policy on every tool call; arguments not documentedInjects downstream credentials; agent never sees themUnified audit trailNot documented
Aembit MCP Identity GatewayToken identifies both the agent and the user; per run not documentedPer MCP server, by client and user; per tool not documentedFetched per request, never given to the agent; per-user OAuth to serversAgent, user, server and policy decisionAembit-managed or self-hosted
PomeriumNot documentedPolicies on tool names, with user, email or group; arguments not documentedInjects upstream OAuth tokens; clients never see themEvery tool call with tool name and parametersNot documented
RunlayerNot documentedRBAC, PBAC or ABAC per server and tool for users, groups or agentsOAuth proxy that stores and refreshes tokensEvery tool call with user, parameters, response; SIEM exportNot documented
Docker MCP GatewayNot documentedServers chosen per profile; per-tool not documentedInjects credentials into server containersBuilt-in logging and call tracingOpen source; Docker Desktop or Engine
IBM ContextForgeNot documentedNot documentedJWT or SSO sign-in; user-scoped OAuth tokensOpenTelemetry tracing, structured logsOpen source (Apache 2.0); self-hosted
Microsoft MCP GatewayNot documentedBasic app-role authorization for servers and toolsEntra ID authenticationApplication Insights telemetryOpen source; Kubernetes (AKS)
ObotNot documentedPer tool, by user, group and agent; arguments not documentedCentral OAuth; tokens stay in the gatewayEvery call with user, agent, arguments, outcomeOpen source (MIT); self-hosted or managed
TrueFoundryNot documentedWhich tools each user can invokePer-user OAuth2, client credentials or static keysRequests attributed to the userVPC, on-prem, air-gapped
Portkey (Prisma AIRS AI Gateway)Not documentedTools enabled or disabled per userInjects OAuth tokens, API keys or headersEvery call with user, parameters, responseNot documented
MintMCP (Mintive)Not documentedRole-based tool sets; destructive tools can be turned offCredentials configured centrallyAudit logs; SIEM exportManaged cloud; on-prem on request

Three tests show whether a gateway decides each tool call

Most gateways look alike at sign-in. Run these during a trial to see how each one behaves after it:

  1. Revoke one agent mid-task. An authorizing gateway refuses that agent's next tool call, and the same person's other agents keep working. With a routing-only gateway, you usually rotate a shared key, which stops every agent that uses it.
  2. Deny one tool for one user. An authorizing gateway removes it from that user's tool list and records any attempt to call it. A routing-only gateway shows everyone the same tools.
  3. Allow the safe tools on a server and deny the destructive ones. For example, let an agent open GitHub pull requests but not merge them or push files. An authorizing gateway blocks the merge and records the denied call. A routing-only gateway can only allow or block the whole server.
What happens to the next tool call when you revoke one agent
REVOKE A SHARED KEYAgent AShared keyStoppedAgent BShared keyStoppedKey revokedEvery agent on the key stopsREVOKE ONE AGENT INSTANCEAgent AOwn identityNext call rejectedAgent BOwn identityKeeps workingA revokedThe user's other agents keep working
With a shared key, revoking access stops every agent on that key. With an identity per agent run, revoking one instance rejects that instance's next call, and the user's other agents keep working.

Security teams judge MCP gateways on identity and access first

A gateway gives security teams one place to manage and audit agent access. Without one, API keys sit on laptops and offboarding means hunting through config files.

How agents reach MCP servers with and without a gateway
AGENTS CONNECT DIRECTLYAGENTS CONNECT THROUGH A GATEWAYClaudeGitHubCursorSalesforceScriptDatabaseAn API key on every laptop, no central viewClaudeGitHubCursorSalesforceScriptDatabaseGatewayOne place to grant, revoke and audit
Each client otherwise connects to each server with its own keys. A gateway routes those connections through one point you control.

When you evaluate gateways, check for:

  • SSO and SCIM. Access follows joiners and leavers in Okta or Entra.
  • An identity per agent. Each agent run is identified and tied to the person it acts for, so you can revoke one agent without cutting off the person.
  • Per-tool rules, down to arguments. A role can read CRM records without seeing the delete tool, and a tool can be limited by what it's called with.
  • No standing credentials. Agents get short-lived tokens instead of long-lived API keys.
  • Audit per call. Each call names the person and the agent, and denied calls are recorded too.
  • Deployment fit. SaaS, VPC, on-prem or open source, depending on your data rules.
  • MCP auth support. OAuth 2.1 and RFC 9728 for remote servers that use authorization.

Per-request authorization closes the gaps a sign-in check leaves

Many gateways check who you are when a session starts and then apply role-based rules to tool calls. That leaves gaps that show up in production:

  • One agent misbehaves, and the only off switch is a shared key that stops every agent using it.
  • A prompt-injected agent calls a tool its role allows, with arguments the delegating user should never use.
  • Several agents share one service credential, so the audit log shows the service account and loses the person behind the call.

Closing these gaps takes authorization on every request, based on the agent and the delegating user, with no standing credentials on the agent. Some gateways include it. If yours doesn't, add an authorization layer next to it.

SecureAuth MCP Authority works downstream of your identity provider. It gives each agent run its own identity tied to the delegating user, checks every tool call against policy down to its arguments, and replaces API keys with short-lived OAuth tokens. Denied calls are recorded, and revoking an agent instance rejects its next call.

Explore MCP Authority by SecureAuth

Frequently Asked Questions

What are the best MCP gateways in 2026?

Enterprise MCP gateways commonly compared in 2026 include Kong AI Gateway, Azure API Management, agentgateway, Cloudflare MCP server portals, Okta Agent Gateway, SecureAuth Agent Authority (MCP Authority), Aembit MCP Identity Gateway, Pomerium, Runlayer, Docker MCP Gateway, IBM ContextForge, Microsoft MCP Gateway, Obot, TrueFoundry, Portkey and MintMCP. The best fit depends on the stack you already run and on whether you need each agent and each tool call authorized, not only routed.

What is an identity-first MCP gateway?

An identity-first MCP gateway authenticates the person, and in some cases the agent, through your identity provider before any tool call, then uses that identity for tool policy, credentials and audit. Examples include SecureAuth Agent Authority, Okta Agent Gateway, Aembit MCP Identity Gateway, Pomerium, Runlayer and Cloudflare MCP server portals. They differ in whether each agent run gets its own identity tied to the user, and in whether policy reaches individual tools and their arguments.

Which MCP gateways are open source?

Docker MCP Gateway, IBM ContextForge (Apache 2.0), Microsoft MCP Gateway, Obot (MIT), agentgateway (a Linux Foundation project) and Agent Router, formerly Envoy AI Gateway (Apache 2.0), are open source. Obot also offers a managed version.

What should an enterprise MCP gateway check on each tool call?

It should check which agent is calling, which person that agent acts for, whether policy allows that tool, and whether the call's arguments are allowed. It should record the decision, denied calls included, and use short-lived tokens rather than standing API keys.

Is SecureAuth Agent Authority an MCP gateway?

Yes. MCP Authority is the MCP gateway path in SecureAuth Agent Authority, one of its three enforcement paths. It works downstream of your existing identity provider, gives each agent run its own identity tied to the delegating user, and checks every tool call against policy down to its arguments. Short-lived OAuth tokens replace API keys, denied calls are recorded, and revoking an agent instance rejects its next call. Agent Authority also covers the path to AI model providers through LLM Authority, which allows only approved models per team, caps spend and strips sensitive data from prompts.

Do I need an MCP gateway if I already have an API gateway?

An API gateway routes HTTP traffic. MCP adds tool discovery and tool calls inside JSON-RPC messages, so you need something that reads those messages to filter tools and log calls. Some API gateway vendors, such as Kong and Azure API Management, now do this.

What does enterprise MCP mean?

Enterprise MCP usually means running MCP across many teams with central access control, SSO, audit and approved server catalogs. The MCP project added an Enterprise-Managed Authorization extension so the corporate IdP can decide which MCP servers each employee can reach.

Does an MCP gateway handle SSO?

Most do for the people signing in. SSO decides who reaches the gateway, so also check that the gateway identifies each agent and authorizes every tool call, not just the session.

More in Agentic AI Security 101

See what your AI agents can reach, and decide what they should.

Agent Authority by SecureAuth gives each AI agent you connect its own identity and checks each action against policy before it runs.