Anthropic extends five agent-building capabilities to Azure-hosted Claude deployments in Microsoft Foundry

Anthropic has extended five agent-building capabilities for Claude models in Microsoft Foundry to deployments hosted on Azure, removing a trade-off between residency requirements and access to the full agentic feature set. The additions—structured outputs, web search, web fetch, an MCP connector, and tool search—were previously available only on "Hosted on Anthropic" deployments within Foundry, according to an announcement from the Microsoft Foundry product team.

Teams with data-handling commitments requiring prompts and completions to remain within Azure previously had to either forgo these features or rebuild equivalents client-side: a retry loop for malformed JSON, a custom search-and-scrape service, a hand-rolled MCP client, and a tool router. The announcement frames all four as "undifferentiated engineering" better owned by the platform.

Structured outputs constrain model decoding with a grammar compiled from a JSON Schema, so generated JSON cannot be malformed—an alternative to the repeated re-prompting loop common in LLM data pipelines. The feature also covers strict tool use, guaranteeing schema-valid tool inputs. Schemas are cached for 24 hours and zero-data-retention processed, with a warning not to place PHI in schemas.

Web search adds citations and, on newer tool versions, dynamic filtering: Claude writes and runs code that filters results before they reach the context window, avoiding wasted input tokens from page boilerplate. Domain restrictions via allowed_domains serve as a security control against prompt injection in the web fetch case.

The MCP connector, still in beta, lets the Messages API connect to remote MCP servers without a client implementation or tool-schema translation layer. It supports granular tool governance—denylists, allowlists, and deferred loading via defer_loading—recommended for anything touching identity, endpoints, or funds. The MCP connector's server exchange is not covered by zero-data-retention processing.

Tool search addresses a documented failure mode: tool-selection accuracy degrades beyond roughly 30–50 tools, and loaded tool definitions can consume thousands of context tokens per turn. It loads only the 3–5 tools a request needs, cutting definition tokens by over 85%, and requires at least one non-deferred tool. One example covers an agent with over 600 tools across nine MCP servers.

Pricing: web search costs $10 per 1,000 searches; web fetch and tool search add no per-call charge. Everything bills as Claude Consumption Units through Azure Marketplace, metered hourly. Features not available on Foundry include the Message Batches API, Admin API, Claude Managed Agents, and server-side fallback.

On Azure-hosted deployments, prompts and completions stay within Azure, while usage metadata and content flagged by Anthropic's safety systems egress to Anthropic. For regulated workloads, that distinction matters, and Anthropic recommends getting both structured outputs' schema caching and the MCP connector's data handling reviewed before production use.