Independent reporting on artificial intelligence.


The Frontier Wire

Comparisons

The 10 best MCP gateways in 2026, compared on security and control

The Model Context Protocol leaves authorization optional and tool safety to the implementer. An MCP gateway is where most teams end up enforcing both. Ten gateways, compared on the controls the specification asks for.

Updated October 5, 2026: expanded from nine to ten gateways with the addition of Obot, and re-checked every entry against current documentation. Microsoft MCP Gateway now requires the 2026-07-28 protocol revision and routes each request independently instead of by session; Kong’s per-tool ACLs date from Gateway 3.13 and its OAuth2 plugin gained token exchange in 3.14; Bifrost’s token exchange now requires a SCIM identity provider.

TL;DR

  • An MCP gateway sits between MCP clients and MCP servers, aggregating tools behind one endpoint and enforcing authentication, per-consumer tool access and logging on every tool call.
  • The MCP specification makes authorization optional, forbids token passthrough and says tools “represent arbitrary code execution”; the useful test for a gateway is how much of that it enforces for you.
  • Bifrost ranks first because one process governs both model calls and tool calls with the same virtual keys, supports inbound OAuth 2.1 and six outbound auth types, and denies MCP tools by default to keys that have not been granted them.
  • Azure API Management, Obot and IBM ContextForge are the strongest alternatives; Kong, Docker, Cloudflare, LiteLLM, Microsoft MCP Gateway and Amazon Bedrock AgentCore Gateway fit narrower deployments built on a platform a team already runs.
  • The current MCP revision (2026-07-28) removes protocol-level sessions and deprecates Dynamic Client Registration, which changes both routing and OAuth requirements for gateways.

An MCP gateway is a proxy that sits between the applications that call tools over the Model Context Protocol and the MCP servers that provide those tools, so that authentication, tool permissions and logging are enforced in one place instead of in every client. Choosing the best MCP gateway comes down to how much of the protocol’s security guidance a product enforces, and whether it also governs the model traffic that decides which tools get called. This comparison ranks ten options, from open-source gateways that handle LLM and MCP traffic in one process to container-based, Kubernetes-native and managed alternatives. Every capability below comes from each vendor’s own documentation or repository as read in early October 2026; none of the products were run under load for this piece.

What an MCP gateway does

An MCP gateway terminates MCP connections from clients, holds connections to many upstream MCP servers, and presents their tools as one catalog. On the way through, it decides which caller may see and call which tool, attaches the right upstream credential, and records the call.

The MCP specification describes three roles: hosts (LLM applications such as an IDE), clients (connectors inside the host) and servers (services that provide tools, resources and prompts). Without a gateway, every host keeps its own list of servers, its own copies of API keys and its own idea of which tools are safe, and nobody holds a shared record of what ran.

A gateway changes that arrangement in four ways:

  • Aggregation: clients point at one endpoint and discover tools from every upstream server behind it.
  • Identity and access: the gateway authenticates the caller and filters the tool list to what that caller is allowed to use.
  • Credential custody: upstream API keys and OAuth tokens live in the gateway, not in each client’s configuration file.
  • Visibility: every tools/call passes one point where it can be logged, rate-limited or blocked.

The servers themselves are a separate choice. The companion guide to the best MCP servers covers which servers are worth connecting in the first place; this piece covers the layer that governs them.

What the MCP specification asks for

The specification’s security guidance is the most useful yardstick for an MCP gateway, because it says plainly which controls are optional and which are forbidden. Authorization is optional, tool safety is left to implementers, and passing a client’s token straight through to a downstream API is prohibited.

The authorization chapter opens with “Authorization is OPTIONAL for MCP implementations.” When an HTTP-based server implements it, the server acts as an OAuth 2.1 resource server, must publish Protected Resource Metadata (RFC 9728) and must validate that tokens were issued for it as the audience; clients send a resource parameter (RFC 8707) and use PKCE (Proof Key for Code Exchange, which stops a stolen authorization code from being redeemed). The chapter also states: “MCP servers MUST NOT accept or transit any other tokens.”

The Security Best Practices document names the attacks that matter for anything in the middle:

Risk named in the specificationWhat it means for a gateway
Confused deputyA proxy using a static OAuth client ID with a third-party API “MUST implement per-client consent” before forwarding.
Token passthroughThe gateway must issue or exchange its own upstream tokens, never forward the client’s token.
SSRF during OAuth discoveryServer-side MCP clients should block private IP ranges and validate redirect targets.
Local MCP server compromiseLocal servers run with the client’s privileges; sandboxing and explicit consent are recommended.
State handle hijackingPossession of a server-minted state handle must never count as authentication.
Scope minimizationBroad scopes widen the blast radius; start with minimal scopes and elevate per operation.

The specification’s overview adds the principle that underlies all of it: “Tools represent arbitrary code execution and must be treated with appropriate caution,” and “Hosts must obtain explicit user consent before invoking any tool.” It also concedes that MCP “cannot enforce these security principles at the protocol level,” which is the gap MCP gateways fill. A broader map of the controls around that gap, from scanners to runtime guards, is in the companion guide to MCP security tools.

What changed in the 2026-07-28 revision

Two changes in the 2026-07-28 revision, which the versioning page still lists as current in October 2026, affect gateway design directly.

First, the revision removes protocol-level sessions, the Mcp-Session-Id header and the initialize handshake. Every request now carries its protocol version and client capabilities, so a gateway can route each request to any healthy server instance. Session-affinity routing matters only for servers that stay on 2025-11-25 or earlier, and many deployed servers will for some time.

Second, the revision deprecates OAuth Dynamic Client Registration (RFC 7591) in favour of Client ID Metadata Documents, in which a client identifies itself with an HTTPS URL that hosts its metadata. Dynamic registration remains for backwards compatibility. It also asks clients to validate the iss parameter (RFC 9207) to defeat mix-up attacks. A gateway that acts as an authorization server for its own MCP endpoint will need to follow both changes.

The operational side of running MCP servers, from tool budgets and naming to approval gates and health checks, is covered in MCP best practices.

How an MCP gateway relates to an AI gateway

An LLM gateway governs traffic to model providers; an MCP gateway governs traffic to tools. The two meet inside the agent loop, where a model proposes a tool call, the tool runs, and the result returns to the model. Running both in one process lets one key carry model and tool permissions together.

The LLM side of that loop is covered in the Frontier Wire explainer on running an AI gateway in front of every model call, and in the distinction between an LLM gateway and an API gateway. For the model-routing products themselves, Maxim’s ranking of production AI gateways scores five of them on overhead, failover, governance and MCP support.

The practical question is whether tool and model permissions live in the same place. When they are separate, a key limited to a cheap model can still drive an agent that calls a database tool with no budget attached. When they are combined, one key can say “this team may use these models, spend this much, and call these twelve tools.” Bifrost, Azure API Management, Obot, LiteLLM and Amazon Bedrock AgentCore Gateway offer versions of the combined model. Teams that want to run the model half on their own infrastructure can compare open-source LLM proxies and gateways built for self-hosting.

MCP gateway evaluation criteria

The criteria below follow the specification’s security guidance and the operational questions that come after it. They are published before the ranking so readers can weight them differently.

CriterionWhat was checkedWhy it matters
Inbound authenticationClient authentication; OAuth 2.1 with RFC 9728 discovery and PKCEThe specification’s authorization model assumes it; clients such as Claude Code expect it.
Outbound credentialsShared keys, admin OAuth, per-user OAuth, token exchangeDelegated credentials avoid shared service accounts and token passthrough.
Tool filtering per consumerDifferent tools per key, user or team; default allow or denyLeast privilege; the specification treats every tool as code execution.
Execution controlAutomatic execution or approvalMaps to the specification’s consent principle.
Audit and observabilityTool calls logged with caller identity; admin changes auditedIncident response and compliance evidence.
DeploymentBinary, container, Kubernetes or managed; licenseWhere tool arguments and results may travel.
LLM gateway in the same productSame keys or identities for model and tool callsOne policy for the whole agent loop.
Token efficiencyFewer tool definitions sent to the modelLarge catalogs inflate input tokens on every turn.
Endpoint coverageVisibility into MCP servers configured on laptopsA server-side gateway cannot see servers that bypass it.

Token efficiency is easy to underrate: every visible tool definition is sent to the model on each turn, the same cost dynamic described in the piece on reasoning-model inference costs. Endpoint coverage is newer as a criterion; only two products in this list document a device-side component.

Top 10 MCP gateways compared at a glance

The table ranks ten MCP gateways on the criteria above. Bifrost is first because it covers every row, including the combined LLM and MCP policy; the others are ordered by how broadly they cover the criteria for a general team.

Rank and gatewayDeployment and licenseInbound authTool filtering per consumerLLM gateway in same product
1. BifrostSelf-hosted binary, Docker or Go SDK; Apache 2.0 core, paid enterprise tierVirtual key headers or built-in OAuth 2.1 server (RFC 9728, PKCE, RFC 8707)Per virtual key, deny by default; Virtual MCP endpoints per slugYes
2. Azure API ManagementAzure-managed, plus self-hosted gateway; classic and v2 tiers, not workspaces; proprietarySubscription keys; OAuth 2.1 with Entra ID; JWT validationTools chosen per server; access per product and subscription, not per toolYes (APIM AI gateway)
3. ObotDocker or Kubernetes; Go; MIT, with Community and Enterprise editionsIdentity-provider login; scoped Obot API keysServer and tool access by user or IdP group; virtual MCP serversYes
4. IBM ContextForgePyPI, containers or Helm; Python; Apache 2.0JWT, OAuth, custom schemes; RBACVirtual servers bundling selected toolsNo (federates MCP, A2A, REST, gRPC)
5. Kong AI GatewayKong Gateway 3.12+; MCP plugins are AI Gateway Enterprise onlyAI MCP OAuth2 plugin as resource serverPer-tool ACLs by consumer and consumer group (3.13+)Yes (Kong AI Gateway)
6. Docker MCP GatewayDocker CLI plugin; Go; MITLocal; OAuth flows for upstream serversPer-profile tool allowlistsNo
7. Cloudflare MCP server portalsManaged, Cloudflare OneCloudflare Access identity policies; service tokensCurated tools and prompts per portalNo (separate AI Gateway product)
8. LiteLLMSelf-hosted proxy; open-source core with enterprise tierLiteLLM keys; OAuth and token exchange to upstream serversPermissions by key, team or organizationYes
9. Microsoft MCP GatewayKubernetes (AKS or local); C#; MITEntra ID bearer tokensRole-based read and write permissionsNo
10. Amazon Bedrock AgentCore GatewayFully managed AWS serviceOAuth or IAM inbound; managed outbound credentialsInterceptors and Cedar policiesYes (model-based routing)

1. Bifrost

Bifrost, the open-source gateway written in Go by Maxim AI, acts as both an MCP client, connecting to upstream servers over STDIO, HTTP or SSE, and an MCP server that exposes the aggregated tools at a single /mcp endpoint. The same process routes model requests to 20+ providers, so one virtual key governs models, budgets and tools. The core is Apache 2.0; clustering, SSO, guardrails and signed audit logs sit in a paid enterprise tier.

The design choice that stands out in the MCP documentation is that “By default, Bifrost does NOT automatically execute tool calls.” On the LLM path, a model’s tool calls come back as suggestions; the application reviews them and calls a separate execute endpoint. That default matches the specification’s consent principle more closely than gateways that run whatever the model asks for.

Authentication in both directions

Inbound, a client can present a virtual key in a header, or connect through a browser flow in which Bifrost acts as an OAuth 2.1 authorization server for /mcp. That flow implements Protected Resource Metadata (RFC 9728), authorization server metadata (RFC 8414), Dynamic Client Registration (RFC 7591), PKCE with S256, and resource indicators (RFC 8707) that bind the token’s audience to /mcp, which lines up with the specification’s authorization chapter. A mode setting chooses whether /mcp accepts headers, OAuth tokens or both, and a request carrying both credential types is rejected.

Outbound, Bifrost supports six auth types for upstream servers: none, static headers, per-user headers, admin OAuth with automatic refresh, per-user OAuth, and token exchange. Per-user types store a credential against the caller’s identity and surface a consent link the first time a tool needs it. Token exchange swaps the caller’s identity-provider token for a short-lived token scoped to the upstream server on each call (RFC 8693 or an equivalent on-behalf-of grant): “The exchanged token is sent upstream; the caller’s original token never is.” That is the specification’s no-passthrough rule implemented as a feature. It is an enterprise feature that requires Bifrost v2.0.0 or later and an enabled SCIM identity provider.

Tool filtering per virtual key

MCP tool filtering is deny-by-default: a virtual key with no MCP configuration sees no MCP tools, except from servers an administrator marks “Allow by Default.” A request header can narrow the allowed tools but never widen them, the allow-list is checked again at execution time, and inactive or expired keys get a 403. Virtual MCPs go one step further: an administrator bundles selected tools from several servers into an endpoint at /mcp/<slug>, reachable only through the keys attached to it. Virtual MCPs are part of the open-source gateway; enterprise adds access profiles that grant them to user groups, data access control and cluster propagation. The Bifrost governance model attaches models, budgets, rate limits and MCP tools to the same keys and access profiles.

Agent mode and code mode

Agent mode lets Bifrost run the tool loop itself, but only for tools listed in tools_to_auto_execute, which must be a subset of the tools the client may execute. Everything else is returned for approval, and the loop stops at a configurable maximum depth (10 by default). Agent mode does not work with streaming responses.

Code mode addresses tool-catalog bloat. Instead of sending every tool definition to the model, it exposes four meta-tools; the model reads tool stubs on demand and writes a short Python (Starlark) script that Bifrost runs in a sandbox, so intermediate results stay out of the context window. In Bifrost’s published benchmark, input tokens fell 58.2 percent with 96 tools across six servers and 92.8 percent with 508 tools across 16 servers, with a 100 percent pass rate in both modes.

Logging, audit and performance

Bifrost keeps LLM log entries that include tool-call arguments and results, and can attach request headers to them as metadata. The MCP gateway page says each tool call is logged with tool name, server, virtual key, latency and the parent LLM request. HMAC-signed audit logs of administrative changes, with export to SIEM formats and S3 or GCS archival, are enterprise features, as are SSO, clustering and content guardrails for PII, secrets and prompt injection. The gateway’s published benchmark reports 11 µs of added overhead per request at 5,000 requests per second on a t3.xlarge.

Enterprise tier: token exchange, group-level access profiles and signed audit logs are part of Bifrost Enterprise.

Best for: organizations that want one control point for the whole agent loop, with tools, models, budgets and credentials governed by the same keys, open-source code they can self-host, and an OAuth implementation that follows the MCP authorization model.

Broad MCP gateway platforms

Three products come closest to Bifrost’s coverage, each built around a different assumption: an existing API management estate, a governed catalog of approved servers, or federation of existing APIs.

2. Azure API Management

Azure API Management (APIM) handles MCP in two ways: it can expose a REST API it already manages as an MCP server, so that “API operations become MCP tools,” or it can sit in front of an existing remote MCP server. Both paths use the same policy engine as the rest of APIM. MCP servers are supported in the classic Developer, Basic, Standard and Premium tiers, in the v2 tiers and on the self-hosted gateway, but not in workspaces. Transports are Streamable HTTP and the older HTTP+SSE, and an existing server placed behind APIM “must conform to MCP version 2025-06-18 or later.”

Inbound, callers authenticate with subscription keys or with OAuth 2.1 through Microsoft Entra ID, validated by policy. Outbound, APIM’s credential manager can inject OAuth tokens for the upstream server under a managed identity. Administrators choose which operations become tools, and a June 2026 update made MCP servers part of APIM products, so access, subscriptions and quotas can be granted to specific consumers. Every MCP request writes an Application Insights record with the tool name, client and authentication type, and Azure API Center can serve as a registry of approved servers. Microsoft describes the AI gateway, “including MCP server capabilities,” as part of APIM rather than a separate offering, so the same instance applies token limits and content safety to model traffic.

Limits: “policies apply to all API operations exposed as tools in the MCP server,” so there are no per-tool rules; servers built from REST APIs expose tools only, not resources or prompts; request headers are forwarded to tool invocations by default, so keeping client tokens away from upstream servers is a policy a team has to write; and response-body logging must be off for streaming to work.

Best for: organizations whose APIs and models already run through Azure API Management and that want MCP access governed in the same control plane.

3. Obot

Obot is an MIT-licensed Go platform that its README describes as shared infrastructure “for connecting AI clients to models and tools, distributing approved MCP servers and skills, managing agent access and credentials.” Its MCP gateway proxies servers that Obot hosts or that run elsewhere, controls “server and tool access by user or identity-provider group,” manages MCP OAuth and both per-user and shared credentials, and can “inspect, reject, or modify MCP requests and responses” through MCP or webhook filters. Version 0.26 replaced composite servers with virtual MCP servers that expose selected tools from several servers.

Three things set Obot apart from the other self-hosted entries. It runs npx, uvx and containerized MCP servers itself as Docker containers or Kubernetes workloads, with domain-based egress rules. It publishes curated, Git-backed MCP and skills catalogs through the standard MCP Registry API, so users install only approved servers. And an LLM gateway in the same platform restricts which models each user can call, keeps provider keys out of clients and records token usage and estimated cost. Audit logs correlate MCP requests and responses, LLM calls and device activity, filterable by user, server, tool or session.

Obot also has an endpoint component. Obot Sentry enrolls devices, inventories installed AI clients and MCP servers, installs hooks for Claude Code, Codex, Cursor and VS Code, and records local tool calls; the README marks device management as beta.

Licensing has a catch worth reading closely. The editions page says the default build supports “up to 100 users and 100 devices” with local, GitHub and Google login; a free Community registration adds Entra, Okta, JumpCloud and Auth0; Enterprise removes the user and device limits and adds support.

Limits: the documentation reviewed does not describe an OAuth token-exchange flow for upstream servers or a deny-by-default rule comparable to Bifrost’s, and larger organizations need a commercial licence despite the MIT code.

Best for: IT and security teams that want an internal catalog of approved MCP servers, hosted server execution and group-based access in one self-hosted product.

4. IBM ContextForge

ContextForge, from IBM, is a Python registry and proxy under Apache 2.0 that federates MCP servers, A2A agents and REST or gRPC APIs behind one endpoint. Its distinguishing feature is virtualization: it wraps REST and gRPC services as MCP tools, and it lets administrators bundle chosen tools into “virtual servers.” The README lists Basic, JWT and custom authentication, user-scoped OAuth tokens, RBAC, an admin UI with air-gapped support, OpenTelemetry, rate limiting and retries, Redis-backed multi-cluster federation and more than 40 plugins. It installs from PyPI, as containers or through Helm charts.

Limits: ContextForge does not route model traffic, so model budgets and tool permissions live in different systems, and the breadth of options means more configuration to secure.

Best for: enterprises with large estates of internal APIs to expose as tools without writing new MCP servers.

MCP gateways built on existing platforms

The next three products add MCP gateway features to something a team may already run: an API gateway, Docker, or a Zero Trust network. Their MCP features are strongest when that platform is already in place.

5. Kong AI Gateway

Kong added MCP through two plugins in Kong Gateway 3.12, both limited to its AI Gateway Enterprise offering. The AI MCP Proxy plugin can pass MCP traffic through to existing servers, convert REST API paths into MCP tools, or aggregate converted tools into one MCP server. From 3.13 it supports per-tool ACLs that allow or deny consumers and consumer groups, with every access attempt written to an audit log; from 3.14 authorization failures return HTTP 403. The companion AI MCP OAuth2 plugin makes Kong the OAuth resource server for MCP traffic, rejects invalid tokens with a 401, and does not forward access tokens upstream by default. Since 3.14 it can exchange the client’s token for a different one before forwarding.

Limits: MCP features require an enterprise contract, and policies are written in Kong’s consumer model rather than as MCP-specific roles.

Best for: organizations that already manage their APIs in Kong and want existing endpoints exposed as governed tools.

6. Docker MCP Gateway

Docker MCP Gateway is an MIT-licensed Go plugin for the Docker CLI and the engine behind Docker Desktop’s MCP Toolkit. Docker’s documentation says it “runs MCP servers in isolated Docker containers with restricted privileges, network access, and resource usage.” Secrets stay in Docker Desktop’s secret store, servers are grouped into profiles that can be shared through OCI registries, tools can be allowlisted per profile, and calls are logged and traced.

Limits: it is designed around a single developer’s machine; it has no multi-tenant identity model or model-side controls.

Best for: developer workstations, where the specification’s warning about local server compromise is most relevant and container isolation is the main control.

7. Cloudflare MCP server portals

Cloudflare’s MCP server portals put multiple MCP servers behind one HTTP endpoint protected by Cloudflare Access, with managed OAuth for users and service tokens for machine-to-machine access. Administrators curate which tools and prompts each portal exposes, can alias tool names, and tools are namespaced by server ID. Context-optimization modes and a code mode, which replaces tool definitions with JavaScript execution and can be optional, on by default or enforced, reduce the definitions sent to the model. Tool calls appear in Access logs, and routing through Cloudflare Gateway adds HTTP logging and DLP (data loss prevention) scanning. Portals speak the stateless 2026-07-28 protocol and fall back to 2025 revisions.

Limits: the documentation caps a portal at 80 servers, does not support stdio-only servers, and states that independent MFA, purpose justification and temporary authentication “will not be enforced” for servers reached through a portal.

Best for: companies already on Cloudflare One that want MCP access tied to their existing identity policies.

MCP gateways tied to a proxy or cloud

The last three entries are capable but narrower: one extends an LLM proxy, one is a Kubernetes reference gateway for Azure, and one is a managed AWS service.

8. LiteLLM

LiteLLM, the Python LLM proxy, includes an MCP gateway supporting Streamable HTTP, SSE and stdio. Permissions to MCP servers can be managed by key, team or organization, upstream authentication covers OAuth with PKCE, client credentials, API keys, AWS SigV4 and RFC 8693 token exchange, and MCP usage feeds its cost tracking. The documentation places SSO, audit logs, spend tracking, multi-team management and guardrails under LiteLLM Enterprise.

Limits: the trade-offs of the proxy itself, covered in the review of LiteLLM alternatives, apply to its MCP gateway too.

Best for: teams already standardized on LiteLLM that want basic MCP permissions under the same keys.

9. Microsoft MCP Gateway

Microsoft MCP Gateway is an MIT-licensed C# reverse proxy and control plane for MCP servers running in Kubernetes. Its current version is a breaking change: it “requires MCP 2026-07-28 clients and adapters” and “does not provide legacy initialization, transport sessions, or protocol downgrade.” Each request is authorized and routed independently, and a tool gateway router can run as several instances without session affinity. Control-plane APIs deploy servers (“adapters”) and register tools; Entra ID bearer tokens authenticate callers, read access goes to the creator, configured roles and mcp.admin, and write access only to the creator and mcp.admin. An agents and sessions subsystem is in preview and opt-in.

Limits: clients and servers that still speak 2025 revisions cannot use it, and it is built around servers it deploys into its own cluster.

Best for: Azure-centric teams that host their own MCP servers on AKS and can move to the current revision.

10. Amazon Bedrock AgentCore Gateway

AgentCore Gateway is a fully managed AWS service that turns OpenAPI, Smithy and Lambda targets into MCP tools, aggregates existing MCP servers into one virtual server, fronts other agents through HTTP targets, and routes inference requests across model providers. It handles inbound authentication and outbound credential injection, offers semantic tool search so agents can find tools without loading every definition, and includes observability and auditing. Fine-grained access control works at gateway, tool, operation and parameter level through request interceptors that read JWT claims, or through Cedar policies on IAM principals; AWS recommends that interceptors deny by default.

Limits: tool-level access is something a team builds in interceptor logic rather than a setting, and the service only runs on AWS.

Best for: teams building agents on AWS that prefer a managed service to operating a proxy.

MCP servers on employee machines

A server-side MCP gateway only governs clients configured to use it. Developers also add MCP servers directly to Claude Desktop, Cursor or a coding agent on their own laptops, and those connections never reach the gateway. The specification’s section on local server compromise describes exactly this risk.

Two products in this list now address it. Bifrost’s answer is Bifrost Edge, an endpoint agent in early access, offered at no cost to existing enterprise customers. The gateway remains the control plane where virtual keys, budgets, guardrails and audit logs are defined; Edge routes AI traffic from chat apps, browser AI, coding agents and MCP servers on each device through it, and can be pushed through device-management tools such as Jamf, Intune or Kandji. For MCP specifically, Edge MCP governance reads the MCP configuration inside supported apps (Claude Code, Claude Desktop, Gemini CLI, OpenCode, Codex and Cursor), builds a fleet-wide inventory, sends newly found servers to an approval queue, and enforces allow or deny decisions on the device, so a denied server cannot be used even by an app that was configured with it earlier. Because traffic passes the gateway, endpoint guardrails such as Gitleaks-based secrets detection apply to prompts leaving the laptop.

Obot Sentry, described above, takes a hook-based approach: it inventories AI clients and MCP servers and records local tool calls alongside gateway activity, with device management in beta. Docker’s container isolation is the nearest control among the remaining products, and only for servers launched through its toolkit. Either endpoint product is something to pilot, not yet a compliance control.

Choosing an MCP gateway by deployment

Most teams can narrow the field by where tool traffic is allowed to go and which platform they already run. The table maps common starting points to the first product to evaluate.

Starting pointFirst MCP gateway to evaluateReason
Agents that call both models and internal tools, self-hostedBifrostOne set of keys for models, budgets and tools; deny-by-default tool access
Existing Azure API estateAzure API ManagementREST APIs exposed as tools, with model and tool traffic on one gateway
Internal catalog of approved servers, hosted for usersObotRegistry, hosted servers and group-based access
Many internal REST or gRPC servicesIBM ContextForgeWraps existing APIs as tools
Existing Kong estateKong AI GatewayPer-tool ACLs on existing consumers
Developer laptops running local serversDocker MCP GatewayContainer isolation per server
Existing Cloudflare Zero TrustCloudflare MCP server portalsAccess policies and DLP already in place
AWS-native, managed onlyAmazon Bedrock AgentCore GatewayNo proxy to operate

For a regulated team, the questions to ask any vendor are the specification’s own: does the gateway validate token audience, does it refuse to pass client tokens upstream, does it deny unknown tools by default, and can it prove who called what. A fifth question is now worth adding: which protocol revisions does it speak, and what happens to clients on the other ones. Microsoft’s gateway answers it by refusing older clients; Cloudflare’s by falling back.

Verdict

Bifrost is the strongest general-purpose MCP gateway in this comparison because it is the only open-source option that pairs a standards-based OAuth 2.1 flow for MCP clients, delegated and per-user credentials for upstream servers, deny-by-default tool access per virtual key and a full LLM gateway in the same process. Code mode addresses the token cost of large tool catalogs, and Bifrost Edge extends the same governance to MCP servers on laptops.

The alternatives are credible within their niches. Azure API Management is the closest peer for teams already on Azure, because one gateway governs model and tool traffic. Obot, the new entry, is the most complete option for teams that want a governed catalog and hosted servers, with the caveat of its user limits. ContextForge suits API-heavy enterprises, Docker suits developer machines, and Kong, Cloudflare, LiteLLM, Microsoft and AWS suit teams committed to those platforms.

Limits of this comparison

Capabilities come from vendor documentation and repositories read in late September 2026 and re-checked in early October 2026; no gateway was deployed, load-tested or penetration-tested, and no vendor benchmark was reproduced. License tiers and supported specification revisions shift between releases, as the Microsoft change shows. Several products were considered and left out, including Google Apigee, MCPJungle and Unla, because this list favours products whose MCP security controls are documented in detail. Tool filtering does not solve prompt injection, since a permitted tool can still return text that redirects the model. Pricing was not compared.

Sources

  1. Model Context Protocol specification, current revision (2026-07-28)
  2. MCP specification: Authorization (2026-07-28)
  3. MCP specification: Security Best Practices (2026-07-28)
  4. MCP specification: Key changes in 2026-07-28
  5. MCP specification: Versioning
  6. Bifrost source repository (GitHub, Apache 2.0)
  7. Bifrost docs: MCP overview and gateway mode
  8. Bifrost docs: MCP gateway authentication
  9. Bifrost docs: MCP authentication types
  10. Bifrost docs: token exchange
  11. Bifrost docs: MCP tool filtering per virtual key
  12. Bifrost docs: Virtual MCPs
  13. Bifrost docs: agent mode
  14. Bifrost docs: code mode and benchmark
  15. Bifrost docs: built-in observability
  16. Bifrost docs: enterprise audit logs
  17. Bifrost docs: benchmarking
  18. Bifrost docs: Bifrost Edge overview
  19. Bifrost docs: Bifrost Edge MCP governance
  20. Bifrost docs: Bifrost Edge security and guardrails
  21. Maxim AI: Bifrost MCP gateway
  22. Maxim AI: Bifrost AI governance
  23. Maxim AI: Bifrost Edge
  24. Maxim AI: Bifrost guardrails
  25. Kong: AI MCP Proxy plugin
  26. Kong: AI MCP OAuth2 plugin
  27. Docker MCP Gateway documentation
  28. Docker MCP Gateway repository (GitHub, MIT)
  29. Cloudflare: MCP server portals
  30. LiteLLM: MCP gateway documentation
  31. Microsoft MCP Gateway repository (GitHub, MIT)
  32. Azure API Management: MCP server support overview
  33. Azure API Management: expose an existing MCP server
  34. Azure API Management: secure access to MCP servers

Questions readers ask

What is an MCP gateway?

An MCP gateway is a proxy that sits between MCP clients, such as coding agents and desktop assistants, and the MCP servers that expose tools. It aggregates many servers behind one endpoint and applies authentication, per-consumer tool access, credential handling and logging to every tool call, so each client does not have to connect to and trust each server on its own.

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

An API gateway routes and secures HTTP endpoints. An MCP gateway understands MCP messages, so it can decide which tools each caller may discover through tools/list and call through tools/call, and it can manage OAuth tokens for upstream servers. Some API platforms, such as Kong, add MCP awareness through plugins rather than a separate product.

Do I need an MCP gateway?

A single developer with one or two local MCP servers usually does not. The case appears when many people or services share MCP servers, when servers hold credentials for company systems, or when security teams need a record of which tools were called and by whom. The MCP specification makes authorization optional, so a gateway is often where it gets enforced.

What is the best open-source MCP gateway?

It depends on the deployment. Bifrost (Apache 2.0) suits teams that want MCP and LLM traffic governed by the same keys. Obot (MIT) suits teams that want a catalog of approved servers and hosted MCP servers. Docker MCP Gateway (MIT) suits developer machines. IBM ContextForge (Apache 2.0) suits teams wrapping REST and gRPC services as tools.

Does an MCP gateway stop prompt injection through tool results?

Not by itself. A gateway limits the damage by restricting which tools an agent can call and by requiring approval before execution, but a tool result that contains instructions still reaches the model. Content guardrails on inputs and outputs, and human approval for tools that write or delete data, are separate controls to add.

More comparisons