What is an MCP gateway?
An MCP gateway is a control point between AI clients and MCP servers. Depending on the implementation, it can authenticate callers, restrict tool access, handle downstream credentials and collect telemetry for traffic routed through it. Evaluate those controls with your actual clients, and check whether a direct server connection can bypass them.
Last verified: November 3, 2026. This guide explains an evaluation model and links to implementation guides. Controls described as requirements must be demonstrated in the proposed deployment.
A successful tool call is the beginning of a trial. The harder questions arrive when the wrong person knows a tool’s name, a credential expires, a server adds an operation or an administrator withdraws approval while clients are still open. A useful gateway makes those cases predictable enough to operate.
Use this page as the entry point for the MCP gateway cluster. It covers where the gateway belongs, when to introduce one and what evidence to request. The linked articles carry the protocol setup and individual operating procedures.
When do you need an MCP gateway?
Introduce a shared gateway when maintaining access policy, credentials and evidence separately in every client or server becomes difficult to control. There is no protocol-defined server count at which a gateway becomes mandatory.
For a pilot with one client and one read-only server, direct authentication and server-side authorization may be sufficient. As more teams and clients share integrations, a central enforcement point can reduce repeated policy work. It also becomes another service to secure, upgrade and recover.
Write down the requirement before choosing the component. “Support employees may search assigned tickets; only leads may update status” is testable. “Centralize AI” leaves the important decision unstated.
Before sizing the service, discover the MCP servers already in use. Count local processes and direct remote connections as well as approved catalog entries. A gateway can govern only the requests that reach it.
Follow one request through the gateway
Suppose an employee asks an assistant to search an assigned support queue. The proposed route is:
AI host and MCP client
-> gateway: authenticate caller and authorize the tool
-> reviewed MCP server: validate operation
-> ticket API: enforce business and record permissions
-> result and correlated evidence back to the host
The MCP architecture defines hosts, clients and servers. A gateway adds an intermediate control point; the protocol does not make every gateway implement the same controls.
Authenticate the inbound request. For protected HTTP services, verify the resource metadata, client registration path and token audience against the MCP authorization specification. Test a missing token and a token issued for a different resource. The Entra ID and Okta setup guide works through the identity-provider side.
Return the permitted tool view. The gateway or server can filter discovery for the caller. That makes the tool surface easier to control, but a client may already know a tool name.
Authorize the invocation. Check the requested tool independently on every call. Include arguments and environment where the requirement needs them. The per-user tool-access guide gives a permission matrix and a direct-call denial test.
Use the intended downstream credential. The identity accepted by the gateway does not automatically establish permissions at the ticket API. Define a supported delegated-user or service-identity flow. The MCP authorization security considerations require audience validation and point to the prohibition on token passthrough.
Reconcile the result. Connect the caller, tool and decision to the destination’s outcome. A request count or latency graph does not prove a particular ticket changed. Use the MCP audit-evidence guide to identify which system supplies each fact.
How does it differ from a registry or another gateway?
A registry records what is available and approved. A gateway acts on requests. Connect the two deliberately: withdrawing a catalog entry should lead to a defined runtime change, with a measured propagation interval.
Our private MCP registry guide covers records, owners and withdrawal. The enterprise registry architecture connects discovery, policy, runtime enforcement and lifecycle management.
An LLM gateway sits on the model-access path. An MCP gateway sits on the tool-access path. One product can cover both, so inspect the controls for each route; see LLM gateway vs agent gateway vs MCP gateway. For a broader protocol-routing comparison, use agent gateway vs API gateway.
Two other distinctions prevent unnecessary rewrites. MCP vs REST API explains how a server can reuse an existing business API. MCP vs function calling explains how the host turns a model’s structured request into an executable tool call. None of those boundaries removes the destination’s business rules.
Open source or managed: compare who operates what
Open source describes software licensing and access to code. Managed describes operating responsibility. A provider can operate an open source gateway, and a customer-hosted deployment may include contracted support.
| Attribute | Self-managed gateway | Provider-managed gateway |
|---|---|---|
| Deployment model | Your team deploys and upgrades the runtime | Provider operates the components specified in the agreement |
| Identity and access control | Your team configures and maintains policy and integration | Responsibilities depend on the service; test the same access matrix |
| Audit logging | Your instrumentation, collection and retention design | Confirm available fields, export, retention and customer access |
| Supported MCP transports | Depends on the installed release and adapters | Depends on the service, edition and supported client versions |
| Hosting | Your chosen infrastructure, including dependencies and egress | Verify the locations of runtime, control plane, credentials and records |
For a project shortlist, use open source MCP gateways compared. It separates Docker, ContextForge, Microsoft and agentgateway by their documented starting points. Teams already using Docker can follow the Docker MCP Gateway alternatives guide to decide which execution responsibilities to keep.
For a managed service, ask which upgrades, recovery work and identity integrations the provider actually owns. Use the AI vendor risk assessment for provider access, retention and exit terms. Hosting and support labels are insufficient to establish the full operating boundary.
Which evaluation criteria should decide the trial?
Require evidence for the workflow your employees will use. Keep the requirement, observed behavior and unresolved gap in separate columns of your trial record.
| Criterion | What to ask | Useful evidence |
|---|---|---|
| Client and protocol compatibility | Which client builds, revisions, transports and registration paths work? | Pinned versions and a successful connection from the actual employee client |
| Per-user tool policy | Can this user search but never update? | Native policy, filtered discovery and a denied direct write call |
| Credential handling | Whose credential reaches the destination? | A sanitized flow and destination-side identity verification |
| Approval and release changes | What happens when a tool or description changes? | Reviewed manifest, release record and withdrawal test |
| Audit coverage | Can we explain an allowed, denied and uncertain call? | Correlated records, field coverage and export/recovery tests |
| Deployment and support | Who patches, recovers and changes this service? | Reproducible deployment, ownership and rollback procedure |
Compatibility includes the request and the client
The current transport specification lists stdio and Streamable HTTP bindings. Products may additionally support older modes or bridges. Test the client-to-gateway and gateway-to-server directions separately, including the protocol revision each side expects.
Keep authentication compatibility in that matrix. A service accepting a manually supplied bearer token has demonstrated a different path from a desktop client completing discovery, registration and login.
Tool approval must survive a server update
A tool’s name can stay the same while its description, schema or side effects change. Store the reviewed release and manifest, then decide which changes require approval. The tool-poisoning change-review guide provides a scoped acceptance procedure.
Audit coverage must survive a failure
Try a denied call, an expired downstream credential and a timeout whose business outcome is uncertain. Check what the gateway observed and what the destination recorded. Keep secret values out of logs and limit payload capture to the evidence requirement.
Assurance needs a defined scope
For a SOC 2 report or another required assurance artifact, verify the covered system, deployment and responsibilities. Use the buyer’s guide on report scope to frame the procurement questions. A hosting label does not establish coverage.
Run a pilot that includes withdrawal and bypass
Choose one read tool, one write tool, two identities and ordinary plus restricted test records. Include the business-system owner so the expected record access is agreed in advance.
- Connect the real client and verify the permitted discovery view.
- Complete the allowed read and inspect the downstream identity.
- Request the forbidden write directly by name; verify no backend change.
- Attempt to retrieve the restricted record.
- Withdraw access while the client remains open; measure the effective interval.
- Try the direct server route from an ordinary employee device.
- Exercise a timeout and inspect the destination outcome before retrying.
- Upgrade and roll back a test release, including its policy and catalog state.
The Entra ID tool-access article turns this into a Microsoft identity rollout. The MCP governance checklist assigns the owners and review obligations around it.
Keep adjacent integration requirements visible. ChatGPT connectors and MCP covers the employee-facing connection; MCP and RAG covers retrieval tools. Instructions such as SKILL.md files need their own release review alongside server permissions.
Where Jarvis fits, and what to verify
ASCENDING builds Jarvis, so apply the same trial requirements to our product. The Jarvis MCP gateway page describes permission-scoped discovery, tool-level access controls and downstream OAuth credential handling. Ask to demonstrate those behaviors with the supplied identities and server.
Review the identity, policy and outcome fields available in the actual deployment. Product documentation and metrics are starting evidence; the trial should establish which records you can collect and how access changes propagate.
Bring the workflow and acceptance matrix to a scoped evaluation. A useful outcome is a recorded decision, an owner for every operating responsibility and an explicit list of gaps to close before production.
References
- MCP architecture
- MCP authorization, revision 2026-07-28
- MCP authorization security considerations
- MCP transport bindings
- Jarvis MCP gateway: documented product scope
FAQ — MCP Gateway Questions Buyers Ask
What is an MCP gateway?
An MCP gateway is a runtime control point between AI clients and MCP servers for traffic routed through it. Authentication, tool-level authorization, credential handling, limits and telemetry depend on the product and configuration. Verify those controls and test direct paths that could bypass the gateway.
Is an MCP gateway the same as an API gateway?
An MCP-aware gateway can apply policy to protocol operations such as tools/list and tools/call. An API gateway may also provide that support. Evaluate the configured behavior: can one identity discover and invoke the permitted tool while a direct request to a forbidden tool is denied?
Do I need both an MCP gateway and an MCP registry?
Use both when you need shared approval and discovery as well as runtime enforcement. The registry holds server metadata and approval state; the gateway checks traffic routed through it. Measure how quickly withdrawal reaches existing clients and direct routes rather than assuming the catalog change takes effect everywhere immediately.
What should regulated industries in the US look for in an enterprise MCP gateway?
Check the documented deployment and data flows, the scope of any independent assurance reports, the available audit evidence and the resource types supported by federation. Audit records should identify the caller, tool, decision and outcome without exposing credentials or unnecessary sensitive payloads. Validate each requirement against your deployment rather than inferring compliance from a hosting label.
Should an enterprise choose an open-source or managed MCP gateway?
Choose based on who should operate the gateway and where requests, credentials and audit records may flow. Open-source, self-managed options put deployment and on-call responsibility on your team; managed services may reduce that work but require review of provider access, data flow, service scope and exit terms. The labels alone do not establish security, compliance or hosting boundaries.

