The Model Context Protocol (MCP) is an open standard for connecting AI models to external systems. Instead of every team hand-writing tool definitions and dispatch code for every internal service, a service exposes an MCP server — a standardized endpoint that advertises its tools and executes calls — and any MCP-capable client can use it. The ecosystem has grown quickly: CRMs, ticketing systems, code hosts, and internal platforms increasingly ship MCP servers, which means "integrate Claude with X" often reduces to "point Claude at X's MCP server."
Without the connector, that pointing is still your job: your application must speak the MCP protocol, list the server's tools, translate them into API tool definitions, and shuttle every tool_use block to the server and back. The MCP connector moves that plumbing server-side. You declare the MCP servers in your API request, and Anthropic's infrastructure handles discovery and the call round trips — Claude uses the remote tools as naturally as built-in ones.
Configuration shape
On the Messages API the connector is a beta feature, enabled with the beta header mcp-client-2025-11-20. A request supplies two coordinated pieces: an mcp_servers list declaring the remote servers, and a matching mcp_toolset entry in the tools array that references a declared server by name — the toolset entry is what actually grants Claude access to that server's tools. This two-part shape mirrors how Anthropic's Managed Agents product declares MCP servers (name plus URL, using the Streamable HTTP transport) and then grants them via an mcp_toolset tool. For the current field-by-field request format, work from the official documentation rather than blog snippets — the connector is in beta and details can shift between versions.
Authentication options
Most enterprise MCP servers require auth, and the pattern across Anthropic's MCP surfaces is OAuth-style bearer tokens presented to the MCP server. Two practical notes from the official docs and ecosystem guidance are worth flagging. First, a service's native REST API key is generally not interchangeable with MCP auth — in Managed Agents, for example, a Notion integration token will not work where an OAuth bearer token is expected. Second, in Managed Agents, credentials live in vaults: stored server-side, matched to servers by URL, auto-refreshed, and never exposed to the execution sandbox — the secret is injected at egress by Anthropic-side proxies. On the Messages API connector, confirm the supported authentication fields in the official documentation for your SDK version before designing your secret-handling flow.
Where the connector runs — and where it doesn't
| Platform | MCP connector |
|---|---|
| Claude API (1P) | Beta |
| Claude Platform on AWS | Beta — same-day parity with 1P |
| Amazon Bedrock | Not available |
| Google Vertex AI | Not available |
| Microsoft Foundry | Beta |
On Bedrock and Vertex AI the absence is workable, just laborious: run an MCP client in your own application, enumerate the server's tools, and expose them to Claude as ordinary client tools via the standard tool_use / tool_result loop. You lose the managed convenience; the protocol still does its job of standardizing the integration.
One scaling tip: MCP servers are the fastest way to accumulate dozens of tool definitions, and selection accuracy degrades past a few dozen loaded tools. Pair the connector with tool search so Claude discovers tools on demand instead of carrying every definition in context.
Where to go next
See Managed Agents vaults for Anthropic's most fully developed MCP credential model, and the feature matrix for beta status across platforms.