Microsoft Foundry Adds Five Claude Agent Capabilities

Microsoft Foundry has expanded its Azure-hosted Claude model deployments with five new capabilities: structured outputs, web search, web fetch, Model Context Protocol (MCP) connectors, and tool search. Previously restricted to Anthropic-hosted deployments, these features are now natively available on Azure. This update allows enterprises in regulated sectors to deploy advanced agentic workflows while ensuring that prompts and completions remain strictly within their Azure network perimeter, such as the US Data Zone Standard.

The release targets common engineering bottlenecks. Structured outputs eliminate malformed JSON errors by constraining generation to a compiled schema. The web search feature, priced at $10 per 1,000 searches, includes versions like web_search_20250305 and web_search_20260209, the latter of which dynamically filters results to save context window space. Web fetch extracts text from URLs, converting a 10 kB page into roughly 2,500 tokens, a 100 kB page into 25,000 tokens, or a 500 kB PDF into 125,000 tokens. Meanwhile, the MCP connector links the Messages API directly to remote servers like Jira or ServiceNow without requiring custom client setups.

To address performance degradation when agents manage more than 30 to 50 tools, the new tool search capability uses regex or BM25 queries to dynamically load only the required tools. This reduces definition tokens by over 85 percent, preventing the context bloat of traditional setups that can consume 55,000 tokens just for tool definitions.

For practitioners, these updates shift the burden of building infrastructure scaffolding to the platform. However, developers must still navigate certain limitations. The Message Batches, Admin, Models, and Compliance APIs, along with Claude Managed Agents and the Advisor tool, remain unavailable on Foundry. Additionally, structured output schemas are cached for 24 hours and should not contain protected health information.