
What is an MCP gateway? How it works and when to use one
An MCP gateway connects AI clients to upstream tools through one access point. Learn how it works, how it differs from a server or proxy, and what to evaluate.
The decision in brief
An MCP gateway is an intermediary that gives AI clients a shared MCP endpoint for reaching upstream tools. Depending on the implementation, it can route requests, check identity, restrict access, require approval and record outcomes. MCP governance is the broader work of deciding who may use those tools and under what rules; a gateway is one place to enforce those decisions.
How an MCP gateway works
Model Context Protocol (MCP) lets applications connect to servers that expose tools, resources and prompts. An AI application hosts an MCP client; an MCP server provides the capabilities that client can use. The protocol does not define a separate gateway role.
A gateway presents an MCP server endpoint to the client and routes requests to upstream MCP servers or connectors. For example, a coding assistant can use one governed endpoint to reach an approved ticket system and document store. The gateway controls that route; the upstream systems still enforce their own permissions.
- Discover: the client asks for available tools with tools/list. A gateway can limit the returned tools to those the requester may use.
- Request: the client sends a tool name and arguments with tools/call. The gateway evaluates its configured controls before forwarding or executing the operation.
- Return: the client receives a result, a refusal or a request for review. Check whether the gateway records the decision separately from the outcome of an attempted action.
MCP gateway vs server, proxy and API gateway
These components can overlap, but they answer different architecture questions. An existing API platform may provide MCP-aware capabilities; verify them rather than assuming you need another service.
MCP security covers risks across the client, server, credentials and connected systems. A gateway can enforce controls on traffic it receives, but forwarding a request does not make it safe. MCP servers still need authorization and input validation.
| Component | Main job | What to check |
|---|---|---|
| MCP server | Expose and implement tools, resources or prompts. | Provider permissions, input validation and tool behavior. |
| MCP proxy | Forward MCP traffic between endpoints. | Which controls it adds beyond forwarding. |
| MCP gateway | Offer a shared MCP endpoint and route to upstream tools. | Requester identity, tool access, approvals and outcome records. |
| API gateway | Manage API traffic and service access. | Whether it understands MCP discovery, invocation and sessions. |
When do you need an MCP gateway?
Consider a gateway when several clients or teams need the same access rules, connection ownership, review of proposed changes or records of tool activity. For example, IT may want to revoke one person’s access across approved connections without editing each client separately. Check that the gateway you evaluate actually supports that control.
A direct connection may be sufficient for a limited use case if the server already provides the required authorization and audit. Read-only access still needs scrutiny: a tool that cannot change data can nevertheless disclose sensitive information.
- Access: can you identify the requester and restrict the connection, tool, and resource?
- Changes: can an upstream tool change without silently expanding access?
- Actions: can you require review before an agent sends a message or changes a record?
- Evidence: can you distinguish permission to execute from a confirmed result?
- Coverage: can users reach the same systems through a route outside these controls?
A request through Durin
Each person connects an approved client through their own authenticated Durin MCP endpoint. A tool call must pass current identity, membership, team, connection, credential, reviewed-tool, resource and inspection checks. Sharing a governed route does not mean sharing a personal identity or credentials.
Reviewed requests are allowed by default after those checks. Administrators can deny all tool access, require independent review before an agent makes changes, or deny those changes. The same policy controls apply in sandbox and production; provider authorization and tool readiness still matter.
If approval is required, the action does not execute. The approvers an approval workflow names, or another active administrator, review the request, and the requester must retry the exact call after approval and before expiry. Approval alone does not start the action.
Durin records the decision before dispatch and the outcome after the attempt. A potentially completed write with no reliable confirmation remains uncertain and is not automatically repeated.
What to test before rollout
Start with one reviewed connection and a low-sensitivity lookup. Confirm an allowed request, a denied request and the resulting records. Then use a supported sandbox action to test independent review, an identical retry, expiry and an uncertain outcome.
Check identity revocation, tool changes, credential ownership, and bypass paths as part of the same review. A gateway cannot govern a request it never receives. Review the vendor’s security evidence and contractual commitments separately.
Where it fits with existing controls
Your identity provider supplies identity and membership information. Existing API controls protect upstream services. MCP Governance connects those controls to tool discovery and invocation; it may be provided by a dedicated gateway or by capabilities in an existing platform.
Choose one connection and a concrete requirement, such as limiting a document store to one team or reviewing an external message. Compare your existing controls with the gateway’s behavior for that request, including what happens when access is revoked or the upstream service fails.