The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools and data sources. Anthropic's MCP connector builds MCP support directly into the Messages API: you list remote MCP servers in your request, and Anthropic's infrastructure handles the connection, tool discovery, and tool calls server-side — no MCP client code in your application. It ships behind the beta header mcp-client-2025-11-20 (the earlier mcp-client-2025-04-04 version is deprecated).
Here is the catch for enterprises on a third-party platform. Anthropic's documentation states the availability plainly: the MCP connector is available on the Claude API, Claude Platform on AWS, and Microsoft Foundry — where it requires a Hosted on Anthropic deployment — and it is not currently available on Amazon Bedrock or Google Cloud.
Why the split falls where it does
The pattern becomes intuitive once you notice what the connector actually is: a server-side capability. When Claude calls an MCP tool through the connector, the outbound connection to your MCP server is made from Anthropic-operated infrastructure. The three surfaces where the connector works are exactly the surfaces where Anthropic runs the serving stack: the first-party API, Claude Platform on AWS (Anthropic-operated, running on AWS), and Foundry's Hosted-on-Anthropic deployment option. Bedrock and Vertex AI serve Claude on cloud-provider-operated infrastructure that does not run Anthropic's server-side tool layer — the same reason code execution and the web fetch tool are absent there.
Google's own documentation for Claude as a partner model confirms the boundary from its side: Claude models on Google Cloud support client tools, but server tools are not supported. On Microsoft Foundry, the split runs through the platform: Claude comes in two hosting versions, "Hosted on Azure" (GA) and "Hosted on Anthropic infrastructure," and only the latter carries the MCP connector and Agent Skills.
What Bedrock and Vertex offer instead
Neither excluded platform leaves you without MCP entirely — they just make it your client's job or their platform's job, rather than the Messages API's job.
| Platform | MCP connector? | Alternative paths |
|---|---|---|
| Claude API (1P) | Yes (beta) | — |
| Claude Platform on AWS | Yes (beta) | — |
| Amazon Bedrock | No | Claude Agent SDK; AWS's Bedrock AgentCore (Gateway converts APIs into MCP-compatible tools and connects to existing MCP servers) |
| Google Vertex AI | No | Claude Agent SDK; your own MCP client with the SDK's local helpers |
| Microsoft Foundry | Only Hosted-on-Anthropic deployments | Microsoft Agent Framework with hosted or local MCP tools; Claude Agent SDK |
On Bedrock, AWS's answer is Amazon Bedrock AgentCore, its own agent platform. AgentCore's Runtime works with Claude and with "popular protocols like MCP and A2A," and AgentCore Gateway turns APIs, Lambda functions, and existing services into MCP-compatible tools. Be aware these are AWS-native concepts: AWS's AgentCore documentation does not mention Anthropic's MCP connector at all, and its "skills" are not Anthropic's Agent Skills. See running Claude in Bedrock AgentCore for the details.
On Vertex AI, there is no Google-managed MCP bridge for Claude on the Messages surface. You either run your own MCP client and pass tools to Claude as ordinary client tools (the client-side MCP helpers in the anthropic SDK exist for exactly this), or you move up a layer to the Claude Agent SDK.
The path that works everywhere: the Agent SDK
The Claude Agent SDK — the library form of the agent loop that powers Claude Code — connects to MCP servers itself, from your process, including local command-based servers the remote connector can never reach. Because the MCP client runs on your infrastructure and the SDK only needs a Messages endpoint behind it, this path is officially documented on all four 3P surfaces via environment variables: CLAUDE_CODE_USE_BEDROCK, CLAUDE_CODE_USE_ANTHROPIC_AWS, CLAUDE_CODE_USE_VERTEX, and CLAUDE_CODE_USE_FOUNDRY. If MCP integration must be portable across clouds, build on the Agent SDK, not the connector.
Where to go next
If the connector is available to you, start with the request shape and tool filtering. On excluded platforms, see the Agent SDK on Bedrock, on Google Cloud, and why Vertex has no server tools.