Microsoft published a quiet but consequential announcement on August 17, 2026: five agentic capabilities that were previously unavailable on Azure-hosted Claude deployments are now live. The release resolves a trade-off that had blocked regulated industries from using Claude agents inside their Azure environments for months.
If you are building on Claude in Microsoft Foundry and you have a data residency requirement, this changes what you can ship.
The Problem It Solves
When Anthropic and Microsoft made Claude generally available in Microsoft Foundry in June 2026, the launch addressed procurement and governance. Enterprises could finally run Claude inside their Azure subscription, using Entra ID authentication, Azure Marketplace billing, and their existing network perimeter controls.
What the launch did not address was agentic capability parity. The five features that turn a model endpoint into a production agent platform (structured output validation, live web search, document retrieval, MCP server connectivity, and large-tool-set management) were only available on Hosted on Anthropic deployments. That version routes prompts and completions through Anthropic’s own infrastructure, outside the Azure boundary.
For most development teams, that was fine. For teams in financial services, healthcare, government, and other regulated sectors operating under data handling commitments that require prompts and completions to remain within Azure (or within the United States under a Data Zone Standard deployment), it was a blocker. Every agentic feature required either violating the commitment or rebuilding the functionality from scratch client-side.
As Microsoft’s announcement described the problem:
Every one of those is undifferentiated engineering. None of it is your product. This release moves all four into the platform and, critically, moves them onto deployments hosted on Azure, so you no longer choose between agentic capability and keeping prompts and completions within Azure.
What Shipped
1. Structured Outputs
Structured outputs constrain the model’s decoding at generation time using a grammar compiled from a JSON Schema. The result is that the output cannot be malformed, not as a retry heuristic or a post-generation validator, but because invalid tokens cannot be selected during generation.
Two modes are available and can be used independently or together:
- JSON outputs via the
output_config.formatparameter: constrains the model’s free-text response to a specific schema. - Strict tool use via
strict: trueon a tool definition: validates tool call parameters against the tool’s input schema.
For enterprise workflows where downstream systems depend on predictable JSON, this eliminates the entire class of failures caused by trailing commas, missing fields, or off-schema keys.
Structured outputs are processed under Azure’s zero data retention (ZDR) terms, with one caveat: schemas are cached for 24 hours. Teams handling PHI or other regulated data should not embed identifiers in schemas.
2. Web Search
Web search adds a single built-in tool that gives Claude access to current information. Add web_search_20260209 (or the latest version) to your tools array and Claude decides when to search, runs as many searches as it needs within your configured max_uses limit, and returns an answer with citations attached to the specific content spans it drew from.
The practical architecture change is that you no longer maintain a crawler, a search index, or a re-ranker. The service handles those. For an agent conducting market research, monitoring regulatory changes, or answering time-sensitive questions against live sources, the reduction in infrastructure is significant.
On Azure-hosted deployments, only the web_search_20250305 tool version is supported. Pricing is approximately $10 per 1,000 searches, billed through Azure Marketplace.
3. Web Fetch
Web fetch retrieves the full text of any URL you provide, or processes a PDF as base64 document content. Where web search discovers current information, web fetch reads a specific document you already know about.
For enterprise use cases, the combination is powerful: a research agent can search for recent regulatory filings, identify the most relevant document, and then fetch and analyze its full text, all within a single multi-turn session without any client-side retrieval infrastructure.
On Azure-hosted deployments, only the web_fetch_20250910 tool version is supported. There is no per-call charge beyond standard token costs.
4. MCP Connector
Model Context Protocol has become the dominant standard for exposing enterprise systems (Jira, ServiceNow, Confluence, internal APIs, databases) to language models. Before this release, teams using MCP with Azure-hosted Claude had to implement a client themselves, managing connection state, session lifecycle, and tool schema translation.
MCP connector removes that layer entirely. You point the Messages API at the URL of your remote MCP server. The Foundry service performs the connection and the tool calls on your behalf. If your internal Claude Code deployment already connects to an MCP server, that server is immediately available to your Foundry agents with no additional integration work.
One security note from the official documentation: MCP connector’s server exchange is not covered by ZDR. Teams with strict data handling requirements should confirm this is acceptable before routing sensitive data through MCP-connected tools.
For tool governance, the connector supports allowed_tools allowlists on the MCP server configuration. Use allowlist-style configurations for anything touching identity systems, financial endpoints, or write operations.
5. Tool Search
Tool search addresses what happens when agent complexity scales. An agent with 10 tools works well with all tools loaded into context. An agent with 300 tools does not: context fills quickly, selection accuracy drops, and costs rise.
Tool search inverts the loading model. Instead of injecting all tool schemas into context at the start of each turn, tool search dynamically retrieves the three to five tools most relevant to the current step. The retrieval uses BM25 (keyword-based matching) or regex matching. Both tool_search_tool_bm25_20251119 and tool_search_tool_regex_20251119 versions are supported on Azure-hosted deployments, along with their legacy aliases.
For enterprise agents that need to span many systems (HR, finance, CRM, ticketing, code repositories), this is the capability that makes a unified agent practical without either capping the tool count artificially or accepting degraded selection accuracy.
The Hosting Decision
Microsoft Foundry now offers two hosting paths with nearly equal capability:
| Hosted on Azure | Hosted on Anthropic | |
|---|---|---|
| Inference location | Anthropic-operated, on Azure infrastructure | Anthropic-operated, on Anthropic infrastructure |
| Data residency | Prompts and completions stay within Azure | Prompts and completions go to Anthropic |
| US Data Zone Standard | Yes | No |
| All five capabilities | Yes | Yes |
| Full model catalog | Opus 5, Opus 4.8, Sonnet 5, Sonnet 4.6, Haiku 4.5 | All Foundry models |
| Recommended for | Most production workloads | Models not yet hosted on Azure |
The practical guidance is straightforward: if your organization has a data handling commitment or a US data residency requirement, use Hosted on Azure. You now get all five capabilities without compromise.
If you need a model not yet available on Azure (Fable 5, for example, is Hosted on Anthropic only for now), use the Hosted on Anthropic path. Both paths bill through Azure Marketplace.
Why This Matters for Enterprise AI Builders
The five capabilities individually are not new. What is new is that they are now available inside the Azure boundary, without requiring custom scaffolding. That changes the build calculus in three concrete ways:
Regulated teams can now start. A healthcare team under HIPAA or a financial services team under their data handling agreement can now build a web-search-backed Claude agent without routing any query traffic outside Azure. The US Data Zone Standard option additionally keeps inference within the United States, which satisfies data localization requirements for many US-regulated contexts.
Agent scaffolding debt shrinks. The pattern Microsoft described (retry loops for malformed JSON, bespoke search and scrape infrastructure, hand-rolled MCP clients, custom tool routers) represents months of engineering work that does not differentiate a product. Moving it into the platform means smaller teams can ship production-grade agents faster.
Internal tooling extends to external agents. The MCP connector creates a direct path from internal developer tooling to customer-facing or cross-team agents. A team whose internal Claude Code agents already connect to a Jira MCP server can expose those same capabilities to a production Foundry agent with a single configuration change.
For enterprise teams already running Claude in Foundry, the capabilities are live now. For teams that have been waiting for Azure-native agentic Claude before committing to a build, the wait is over.
The broader context is worth noting. This release arrives as Microsoft continues to expand its Anthropic relationship alongside its established OpenAI partnership. The explicit promise when the expanded Anthropic relationship was announced in late 2025 was that Azure customers would get access to “Claude-specific capabilities.” Five agentic capabilities on Azure-hosted deployments, published August 17, is the follow-through on that promise.
Enterprise AI architects evaluating their model-and-platform strategy now have a cleaner answer: Claude on Azure, with full agentic capability, inside your governance perimeter.
Internal Links
For more context on how enterprise teams are choosing between agentic platforms, see our analysis of how Claude Cowork and ChatGPT Work compare as enterprise AI workspaces. For teams designing the governance layer around multi-agent deployments, the Anthropic multi-agent turf war research from August 18 is directly relevant to why isolation and kill switches matter before you connect agents to production systems. To see how governance is being standardized across the broader enterprise AI stack, the Snowflake Cortex AI Gateway covers the complementary data-side control plane.