Five agent tools reach Claude deployments hosted on Azure

Structured outputs, web search, web fetch, the MCP connector and tool search now run on Claude deployments hosted on Azure in Microsoft Foundry, five capabilities that previously worked only on Hosted on Anthropic deployments. Microsoft detailed the rollout in its Foundry developer blog.

At a glance

• On Hosted on Azure deployments, only usage metadata and content flagged by Anthropic's safety systems leaves Azure, and a US Data Zone Standard deployment keeps inference within the United States.
• Tool search loads roughly three to five tools per request rather than every definition, cutting definition tokens by over 85%; web search bills at $10 per 1,000 searches, fetch adds no per-call charge.
• Claude models went generally available in Foundry in June 2026, but the agentic feature set ran only on Hosted on Anthropic deployments, forcing teams with Azure data-handling commitments to rebuild it client-side.

The trade-off this removes is procurement as much as engineering. Regulated teams that committed to keeping prompts and completions inside Azure had been paying twice for search, fetch, MCP clients and tool routing: once in a platform feature set they could not use, and once in code they maintained themselves. Moving that scaffolding onto Azure deployments reads as an effort aimed at that group rather than at new capability.

Structured outputs constrain decoding with a grammar compiled from the JSON Schema

Structured outputs restrict generation itself: decoding is constrained by a grammar compiled from the supplied JSON Schema, so a malformed response is not possible. The release covers two independent pieces, JSON outputs set through output_config.format, which governs the shape of the response text, and strict tool use, set with strict: true on a tool, which guarantees schema-valid tool inputs.

The Anthropic Foundry SDKs expose it through messages.parse, with Pydantic models in Python and Zod schemas in TypeScript, returning an already validated object. Structured outputs are ZDR-processed, though schemas are cached for 24 hours, which is why the guidance is to keep PHI out of schema definitions.

Web search from version web_search_20260209 filters results in code before they reach context

Web search is added as a single tool entry, and Claude decides when to run it within the max_uses limit, returning citations for the spans it used. Version web_search_20250305 loads every result into context in full; from web_search_20260209 the model instead writes and runs code that filters results before they reach context.

On those versions allowed_callers defaults to the code execution tool, which Foundry provisions automatically at no charge beyond token costs; models without programmatic tool calling must set it to direct or the request returns 400. Usage bills as Claude Consumption Units through Azure Marketplace, metered hourly and invoiced monthly in arrears.

Web fetch reads a URL and returns page text or, for PDFs, base64 document content handled like an attached file; citations are off by default there and max_content_tokens truncates oversized text. The post budgets a 10 kB page at roughly 2,500 tokens, a 100 kB documentation page at 25,000 and a 500 kB PDF at 125,000.

Tool selection accuracy degrades past roughly 30–50 available tools

A setup spanning GitHub, Slack, Sentry, Grafana and Splunk consumes roughly 55,000 tokens in tool definitions on every turn before any work is done, and selection accuracy falls off once more than roughly 30–50 tools are available. Tool search leaves the catalogue out of context and searches it, in a regex variant capped at 200 characters or a BM25 variant capped at 500.

defer_loading controls what enters the context window, not what is transmitted: the full tools array still travels with every request, and deferring every tool returns a 400. The MCP connector, in beta on Foundry under the mcp-client-2025-11-20 flag, declares servers in mcp_servers and per-server policy in mcp_toolset entries, with denylist, allowlist and deferred-loading configurations.

The APIs Foundry does not carry

Several surfaces remain outside Foundry: the Message Batches API, Admin API, Models API and Compliance API, along with Claude Managed Agents, server-side fallback and the Advisor tool. Foundry also does not surface Anthropic's rate-limit headers, so exponential backoff has to be implemented client-side, and Entra ID tokens expire after about an hour. Anthropic recommends at least 30-day rolling retention for request logs routed to Azure Monitor.