Microsoft Foundry offers Claude models in two hosting versions, and the difference is more than an infrastructure footnote. Hosted on Azure runs Claude on Azure infrastructure end-to-end and is generally available, currently with claude-opus-4-8, claude-sonnet-5, and claude-haiku-4-5. Hosted on Anthropic infrastructure runs on Anthropic's infrastructure outside Azure, is in preview, and carries a broader model list — including claude-fable-5. Microsoft's docs note the Hosted-on-Anthropic version "supports more APIs" and defer the full list to Anthropic's documentation. For agent builders, three of those APIs are the whole story.
The three features gated on Hosted-on-Anthropic
Anthropic's own documentation is explicit about each:
- MCP connector — connect remote MCP servers directly from the Messages API. "On Microsoft Foundry, the MCP connector requires a Hosted on Anthropic deployment."
- Agent Skills — pre-built skills (pptx, xlsx, docx, pdf) and custom skills on the Messages API. Available on Foundry, but again only with a Hosted-on-Anthropic deployment; Anthropic's docs add that "Claude Platform on AWS and Microsoft Foundry inherit the same Skills behavior as the Claude API."
- Code execution tool — Anthropic's server-side sandbox, which Skills also depend on. Its availability beyond the first-party API is Claude Platform on AWS and Microsoft Foundry, requiring a Hosted-on-Anthropic deployment.
Because Skills run inside the code execution container, these three features travel together: pick Hosted on Azure and you lose the whole cluster; pick Hosted on Anthropic and all three unlock, along with Fable 5 in preview.
What Hosted on Azure does well
Don't read the gating as "Hosted on Azure is limited to bare chat." Microsoft's capability documentation (updated June 2026) lists a substantial GA surface for Azure-hosted Claude: the Messages API with citations (including search-result content blocks), PDF support, prompt caching, extended thinking, and the 1M-token context window — with only mid-conversation system messages and token budgets flagged as beta. Notably, context editing (automatic clearing of old tool results and thinking blocks, via the context-management-2025-06-27 header) and fine-grained tool streaming are listed as available on both hosting modes. So a self-managed agent loop — your code executing client-side tools, with context editing keeping long sessions lean — works fine entirely on Azure infrastructure.
Both modes expose the same endpoint shape, https://<resource>.services.ai.azure.com/anthropic/v1/messages, callable from the standard anthropic Python package:
from anthropic import AnthropicFoundry
client = AnthropicFoundry(api_key="...", resource="my-foundry-resource")
response = client.messages.create(
model="claude-sonnet-5", # GA on Hosted on Azure
max_tokens=1024,
messages=[{"role": "user", "content": "Summarize this contract."}],
)
print(response.content)Decision table
| Question | Hosted on Azure (GA) | Hosted on Anthropic (preview) |
|---|---|---|
| Where does inference run? | Azure infrastructure end-to-end | Anthropic's infrastructure, outside Azure |
| Models | Opus 4.8, Sonnet 5, Haiku 4.5 | Broader list, incl. Fable 5 (preview) |
| MCP connector / Agent Skills / code execution | No | Yes |
| Context editing, fine-grained tool streaming | Yes | Yes |
| Release status | GA | Preview |
Agent paths on Foundry
Microsoft's Claude concepts page lists exactly two agent-building paths: the Microsoft Agent Framework and the Claude Agent SDK. The Agent Framework's Python client supports function tools, hosted and local MCP tools, hosted web search, and the code interpreter ("Required for Anthropic Skills") — Microsoft's own example runs Anthropic's pptx skill through it — while the .NET client currently supports function and MCP tools but not code interpreter or web search. The Claude Agent SDK targets Foundry with CLAUDE_CODE_USE_FOUNDRY=1; see running the Agent SDK on Foundry. Foundry Agent Service (the portal-managed threads-and-runs pattern) is not documented as supporting Claude.
Where to go next
For the MCP connector's availability picture across all platforms, read the MCP connector platform split; for the Skills request shape, see Agent Skills on the Messages API. The feature matrix summarizes Foundry against the other three platforms.