Somewhere in your organization, there is a compliance requirement that says prompts and completions stay inside Azure. One sentence, buried in a data handling policy nobody rereads. And until this week, that sentence was quietly crippling your Claude agents.

No web search. No document fetch. No MCP connector. No tool routing at scale. Not unless you rebuilt every piece of it yourself and maintained it forever.

That just changed, and it changes how we architect agents in a Microsoft-focused enterprise.

YouTube player

Why the trade-off existed

Claude models became generally available in Microsoft Foundry in June, and they come with two hosting options you can choose when you create the deployment: hosted on Azure or hosted on Anthropic.

The catch was that the full agentic feature set only ran on the hosted on Anthropic option. Teams with a data boundary commitment were stuck with a bad choice. Go off Azure and break the commitment, or keep the boundary and hand-build search, fetch, MCP, and tool routing on the client side.

Microsoft has now made five capabilities available on deployments hosted on Azure. Prompts and completions remain within Azure, with usage metadata and safety-flagged content as documented exceptions. Review the data terms for your specific workload before you ship.

The five capabilities

Structured outputs. You define a schema and Claude’s decoding is constrained by a grammar compiled from it. The output cannot be malformed. For a nightly batch job feeding Azure SQL or a Logic Apps workflow, that means no retry loops and no dead letter queue triage on Monday morning.

Web search. Claude decides when it needs current information, searches, and returns citations tied to specific spans. You can restrict it to trusted domains, which matters when the answer has to come from a regulator and not from a blog paraphrasing that regulator incorrectly.

Web fetch. Search finds the source. Fetch reads the full page or PDF. Together they give an agent both breadth and depth for vendor security reviews and technical documentation. One tip: citations are off by default for fetch, so turn them on.

MCP connector. Model Context Protocol is the standard adapter between Claude and systems like Jira, ServiceNow, and Confluence. Foundry handles the remote connection, and you govern the tools, allowing the read-only ones and blocking the destructive ones. If you already run MCP servers for an internal Claude Code deployment, you are pointing a new consumer at work that is already done. One caveat: the connector is still in beta on Foundry.

Tool search. This one comes with numbers. Per Microsoft, tool selection accuracy can degrade past roughly 30 to 50 tools, and a modest multi-server setup can burn 55,000 tokens on tool definitions before the model does any work. Tool search loads only the three to five tools each task actually needs, cutting definition tokens by over 85 percent.

What is still ours to build

Permissions. Prompt injection defense. Cost controls. Logging. Failure handling. This is not a magic agent button.

But it is considerably less undifferentiated engineering. Microsoft moved the common scaffolding into Foundry, which leaves us free to build the workflows and business logic that actually make our products valuable. After 25 years of integration work, I can tell you that is the trade every architect wants.