News
Microsoft Previews Cloud-Hosted Foundry MCP Server for AI Agent Development
Microsoft is previewing Foundry MCP Server, a fully cloud-hosted implementation of the Model Context Protocol (MCP) for Microsoft Foundry.
MCP is a standard for letting AI agents connect to apps, data, and systems through a consistent, secure interface. Microsoft says the new server delivers curated MCP tools that enable read and write access across Foundry features without requiring agents to call backend APIs directly.
The company previously released an experimental local MCP server for Foundry at Build 2025. With the Ignite 2025 update, the MCP server now runs in the cloud, offering public endpoints and removing the need for developers to host and manage a local server. Microsoft positions this as simplifying setup, speeding integration, and improving reliability for teams building or operating agents.
[Click on image for larger view.] Foundry Ecosystem (source: Microsoft).
Microsoft's post said Foundry MCP Server is reachable through a public endpoint, but that link isn't working at the time of this writing. It supports OAuth 2.0 authentication using Microsoft Entra ID and on-behalf-of (OBO) tokens so the server can act with user-scoped permissions. Tenant administrators can control token retrieval using Azure Policy. Microsoft also says the service is designed for cloud-scale reliability and security.
Microsoft frames the server as part of its broader Foundry agent strategy. At Ignite, the company introduced Foundry Tools as a hub for discovering, connecting, and managing public and private MCP tools, and highlighted compatibility with more than 1,400 business systems. Foundry Agent Service was also updated to support building and connecting agents to those tools. The new MCP Server is intended to be the managed, hosted bridge that agents use to interact with Foundry capabilities through MCP.
The preview exposes tools grouped into several common scenarios. For agent management, MCP tools let developers create, update, clone, list, inspect, and delete agents, including configuring models, instructions, toolsets, temperature, and safety settings. For evaluations, tools support registering datasets, inspecting dataset versions, starting evaluation runs with built-in evaluators, listing evaluation groups and runs, and producing comparison insights between baseline and treatment runs.
Model exploration and governance is another focus. MCP tools can list catalog models, retrieve benchmark overviews and subsets, find similar models based on behavior and benchmark results, and generate switch recommendations based on quality, cost, latency, safety, or throughput goals. Developers can also fetch detailed model metadata and example code snippets. For deployment lifecycle management, tools allow creating or updating model deployments, listing deployments, deleting deployments, and retrieving deprecation timelines and migration guidance.
Operational monitoring is covered through tools that return deployment-level monitoring metrics such as request volume, latency, quota usage, and service indicators, as well as subscription-level quota and usage summaries by region. Microsoft's examples emphasize using these tools in a single conversational flow so an agent can explore options, take action, and then validate outcomes without leaving the chat context.
The announcement gives two end-to-end scenarios to show how tool chaining works. In a build-from-scratch path, an agent can help select a model using catalog and benchmark tools, deploy it if needed, create the agent, run quality and safety evaluations on a dataset, and compare new results to older agents. In a production-optimization path, an agent can inspect monitoring and quota headroom, identify potential replacement models, benchmark candidates, update the deployment, and optionally retire the prior endpoint.
Developers can connect to Foundry MCP Server from Visual Studio Code using GitHub Copilot in agent mode, or from Visual Studio 2026 Insiders, where Microsoft has added MCP server support. The cloud-hosted design is meant to make these connections more lightweight, with the server available as soon as it is added to the IDE's MCP configuration. Within Microsoft Foundry, developers can also attach the MCP Server to agents through the Tools menu and create a tool connection with a single action.
On security, Microsoft says Foundry MCP Server accepts only Entra ID tokens scoped to the MCP endpoint, and every action runs under the signed-in user's Azure role-based access control permissions. This means agents cannot perform operations beyond the user's rights. Activity is logged for auditability, and administrators can apply Conditional Access policies through Azure Policy to manage MCP usage.
Microsoft is positioning the preview as a way to reduce friction for teams building agents that rely on Foundry's model catalog, evaluation services, deployment controls, and monitoring. By hosting the MCP Server in the cloud and curating Foundry-centric tools behind a standard protocol, the company aims to make agent workflows more interoperable across tools and faster to stand up in common developer environments.
About the Author
David Ramel is an editor and writer at Converge 360.