In practice, the integration splits the work. Agent 365 holds the list of approved MCP tools and makes the access decisions, and the gateway enforces them on each tool call. One warning before the details: connecting the two does not automatically send every agent's traffic through the gateway.
The two halves: Agent 365 and the gateway
Microsoft's AI Gateway portal puts it simply: Agent 365 decides which tools each agent may use. The gateway applies those decisions on every call.
The release notes for Azure AI Gateway describe the integration this way:
- It lets you centralize MCP inventory and access decisions in Agent 365, with gateway enforcement on tool traffic.
- Admins can block servers or selected tools without reconfiguring each client.
- Individual tool controls require a static allowed-tool list.
- Runtime signals support auditing and investigation.
Microsoft Learn's integration guide, "Govern tools with Microsoft Agent 365 and Azure API Management," says Agent 365 discovers Azure resources of type Microsoft.ApiManagement in the tenant. It then pulls in the MCP servers and tools those resources expose, along with metadata such as endpoints, access scopes and associated service identities.
Why it matters: in many organizations, MCP servers have multiplied faster than anyone has catalogued them. This integration gives the identity and security side of IT one inventory and one place to block servers or tools, while enforcement stays on the gateway that already carries the traffic.
Related launch: Azure AI Gateway is now its own product
Azure AI Gateway reached general availability on the same day, and the two announcements are easy to mix up. According to Microsoft's AI Gateway GA blog post, the service ran in public preview as the AI Gateway tier of APIM and is now a standalone, fully managed Azure service. Microsoft says it can be used on its own with any app or agent, or alongside Microsoft Foundry and Agent 365.
The preview was much smaller. When Microsoft launched it, the AI Gateway tier was available at no cost in public preview in East US 2 and Sweden Central. At GA, the release notes list 17 regions: Canada East, Central US, East US, East US 2, East US 2 EUAP, North Central US, South Central US, West US, West US 2 in the Americas; France Central, North Europe, Sweden Central, UK South, West Europe in Europe; and Australia East, Japan East, Southeast Asia in Asia Pacific.
Microsoft's GA post also gives pricing: a free tier, then per-use charges, with paid billing starting January 1, 2027. Above 10,000 monthly requests, each gateway costs $500 per month and includes 10 million requests. Beyond that, the charge is $20 per additional million requests. Prices are in US dollars, and Microsoft says they may vary by region.
The wording on the free allowance doesn't match across Microsoft's own pages. The GA blog says 10,000 free requests per gateway each month. The APIM Ignite blog and the AI Gateway portal both say 10K free gateway requests per Azure subscription each month. If you plan to run several gateways in one subscription, get that clarified before you budget.
The regions and pricing above apply to Azure AI Gateway itself. Microsoft has not published separate pricing or a region list for the Agent 365 integration. Agent 365 has its own licensing: an Azure architecture post states that Microsoft Agent 365 is generally available and licensed per user.
Which tier supports which feature
The Learn integration guide includes a capability table. Treat it with some caution, because the page is still titled "(preview)" and lists these boundaries under a heading called "Preview limitations."
| Capability | APIM v2 tiers | AI Gateway tier |
|---|---|---|
| Centralized MCP server inventory | ✔️ | ✔️ |
| Block or unblock MCP servers | ✔️ | ✔️ |
| Tool-level blocking | — | ✔️ |
| Tool activity observability | — | ✔️ |
| Tool threat protection | — | ✔️ |
The same guide says APIM v1 (classic) tiers aren't supported. It also says tool-level governance isn't available for MCP servers with dynamic allowed-tool lists. In that case, Agent 365 can't read the tool list, and Microsoft points admins to the AI Gateway portal to block tools there.
Bottom line: v2 tiers give you inventory and server-level on/off switches. Blocking individual tools requires the AI Gateway tier.
Setup: turning it on in the Microsoft 365 admin center
The steps below come from Microsoft Learn's "Manage Tools Gateway" and integration pages.
Prerequisites:
- A Microsoft Agent 365 tenant
- A supported APIM instance or Azure AI Gateway. Microsoft recommends the AI Gateway tier for the fullest integration.
- One or more MCP servers registered with the gateway, each with a static allowed-tool list if you want tool-level control
- An admin who can grant tenant-wide consent. You must have the Global Administrator role to connect Azure AI Gateway. The AI Administrator role doesn't support this operation.
Steps:
- Sign in to the Microsoft 365 admin center as a Global Administrator.
- Go to Agents > Settings, then select Gateways.
- Next to Microsoft Azure, turn on the toggle to connect an Azure API Management service or Azure AI Gateway.
- Review and accept the requested permissions. Microsoft says consent is needed once per tenant and covers every MCP server in the connected gateway.
- Go to Agents > Tools and refresh. Microsoft notes discovery can take a few minutes. Servers that came from the gateway show AI Gateway as the publisher.
What consent grants: the Learn guide says the admin assigns the "A365 MCP Gateway Policy Operator" Azure role at tenant scope. That lets the Agent 365 service identity read APIM resources, read and write API-scoped policies, and discover supported resources in existing and future subscriptions. After consent, governance actions run under Agent 365's app-only identity, so individual Agent 365 admins don't need their own Azure permissions.
That is a broad grant, and it covers subscriptions you haven't created yet. Security teams should log it, review it like any other tenant-wide service permission, and make sure APIM owners know that a Microsoft-managed identity may now insert policies into their APIs.
What blocking looks like on the wire
The Learn guide documents exactly how blocks behave, which helps when you're debugging an agent that has suddenly stopped working:
- Blocking a whole server: Agent 365 adds a managed
return-responsepolicy at the start of the API's inbound policy. Callers get HTTP 403 with the error codeAccessBlockedByGateway. Unblocking removes the policy. - Blocking a single tool: the call returns HTTP 200, but the JSON-RPC result carries
isError: true, a message telling the agent not to retry, and the denial codeToolNotAvailable. Microsoft designed it this way so agents understand the block and stop retrying.
This matters for monitoring. A blocked tool doesn't show up as an HTTP error, so if your dashboards only alert on 4xx and 5xx responses, they won't catch it. You'll need to parse the response body or use the gateway's tool activity telemetry.
The biggest caveat: agents can bypass the gateway
The Learn guide states directly that Agent 365 does not automatically route agents through a customer-hosted gateway. Developers have to point each agent at the gateway's MCP endpoint.
Microsoft's admin page adds that governance covers MCP servers registered with the connected gateway, and admins shouldn't assume it covers every MCP integration or calls made outside that gateway. If an agent can reach the MCP server directly, your carefully configured blocks don't apply to it.
Microsoft's advice is to restrict MCP server connectivity so the gateway is the only permitted caller, or to use private networking or network-level access controls where supported. Treat this as a required step, not a nice-to-have. Without it, the governance applies only to agents that are already using the gateway.
Threat protection is still in preview
Tool threat protection works alongside the integration. Microsoft Defender and Microsoft Purview inspect tool requests and/or results for malicious content, prompt injection and sensitive-data exposure. Microsoft's GA blog labels it preview, and so does the Learn configuration page.
Requirements, according to Learn:
- A Defender or Purview license
- An MCP server on AI Gateway
- Permission to manage gateway policies
- Callers that send a delegated Microsoft Entra ID user token. App-only service-to-service calls aren't supported, because the evaluation relies on an on-behalf-of token exchange.
Three practical notes from the same documentation:
- You can apply only one policy per MCP server, and it covers all of that server's tools.
- Enabling checks both before and after invocation can add two evaluation round trips to each call.
- A denial after the tool has run withholds the result but cannot undo what the tool already did. Use pre-invocation checks for tools that delete, transfer or send things.
Microsoft's docs still say "preview"
Several Microsoft Learn pages hadn't caught up with the GA announcement at the time of writing:
- The integration guide is titled "Govern tools with Microsoft Agent 365 and Azure API Management (preview)."
- The Manage Tools Gateway page still says the feature is in preview and not meant for production.
- The threat protection page still describes AI Gateway as preview, available only in East US 2 and Sweden Central.
The Azure Updates entry, the Ignite blog posts and the AI Gateway release notes all say GA. Documentation usually lags announcements, but check the tier limits and supported scenarios against the updated pages before you write them into a production design.
What admins should do now
- Inventory first. Connect the gateway and see which MCP servers appear. Expect a few you didn't know about.
- Check your tiers. v1 APIM instances are out. If you want per-tool control, plan for AI Gateway.
- Switch MCP servers to static allowed-tool lists where you need tool-level governance.
- Close the bypass. Lock MCP backends so only the gateway can call them.
- Review the consent grant like any tenant-wide service principal permission.
- Update monitoring so it catches tool blocks that come back as HTTP 200.
- Pilot threat protection carefully. It's preview and requires delegated tokens, which may not suit headless agents.
Together, these launches give Microsoft a full path from agent catalog to network enforcement for MCP tool traffic. The integration only covers agents that actually connect through the gateway, so the job now is making sure they do.
References
- (Launched) Generally Available: Microsoft Agent 365 integration with Azure API Management and Azure AI Gateway Azure Updates · 2026-11-17T20:00:32Z
- Govern tools with Microsoft Agent 365 and Azure API Management (preview) | Microsoft Learn learn.microsoft.com
- Azure AI Gateway is now generally available: Ship faster, stay in control | Microsoft Community Hub techcommunity.microsoft.com