Microsoft has added five new capabilities to Anthropic's Claude models running on Azure through Microsoft Foundry, closing the gap between a raw model endpoint and a production-ready agent platform. Structured outputs, web search, web fetch, an MCP connector, and tool search are now available on deployments hosted on Azure infrastructure.
The features arrive roughly a month after Claude became generally available in Foundry on Azure-hosted deployments, which gave enterprises Entra ID authentication, Azure Marketplace billing, and the assurance that prompts and completions stay within their Azure subscription. The new tooling addresses the scaffolding teams typically rebuild after shipping a first Claude feature: retry loops for malformed JSON, custom crawlers for web search, hand-rolled Model Context clients for internal systems, and tool routers to prevent models from picking the wrong function among hundreds.
What's new
The five capabilities share a common theme: they move undifferentiated engineering into the platform while keeping inference on Azure-hosted deployments. Previously, agentic features were only available on "Hosted on Anthropic" deployments, forcing a trade-off between data-residency commitments and agent functionality.
- Structured outputs constrain generation with a grammar compiled from a JSON Schema, eliminating malformed JSON. Two modes are supported: JSON outputs via output_config.format for response shape, and strict tool use (strict: true) for schema-valid function arguments.
- Web search (version web_search_20250305 and newer web_search_20260318) lets Claude decide when to search, run multiple queries, and return answers with inline citations. The newer version adds dynamic filtering via code execution to reduce context-window bloat from irrelevant result boilerplate.
- Web fetch (version web_fetch_20250910) retrieves full page text or base64-encoded PDF content from a given URL, processed like a directly attached document.
- MCP connector allows Claude to call tools exposed through Model Context servers — Jira, ServiceNow, Confluence, and internal APIs — without a custom client.
- Tool search helps the model select the right tool from large catalogs (hundreds of functions) by searching descriptions before invoking.
All five are available on both Global Standard and US Data Zone Standard deployments. On Azure-hosted deployments, only usage metadata and content flagged by Anthropic's safety systems egress to Anthropic; prompts and completions remain in Azure.
Why it matters
For regulated industries, the combination of US data-zone residency and agentic tooling removes a twelve-month-old dilemma: teams no longer choose between keeping data in-region and giving models web search, MCP access, and grammar-constrained outputs. A nightly regulatory-monitoring job can now run on a US Data Zone deployment, restrict search to a curated list of regulator domains (allowed_domains), and emit validated JSON that feeds directly into downstream Azure SQL and Logic Apps workflows.
The web search dynamic filtering is a token-economics shift. Basic search loads entire result pages — navigation chrome, boilerplate, unrelated paragraphs — into context. The web_search_20260318 version runs filtering inside the code execution tool (provisioned automatically, no extra charge beyond tokens), keeping only relevant snippets. Teams can force direct calls with allowed_callers: ["direct"] for models that don't support programmatic tool calling.
Our take
The real unlock isn't any single feature — it's that the entire agent scaffolding now runs on Azure-hosted deployments with the same governance, billing, and networking controls enterprises already use. That turns "Claude on Azure" from a procurement win into a platform decision.
