MCP connector: mcp_servers and mcp_toolset

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your team connects a calendar MCP server to Claude Opus 5 (claude-opus-5) through the MCP connector. The server exposes eleven tools, and only search_events and list_events may be reachable from the assistant. Which approach should you take?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: A is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   The answerA.Set default_config.enabled to false on the toolset and enable the two in configs.The allowlist pattern sets default_config.enabled to false on the mcp_toolset entry, then re-enables the wanted tools in configs. Tool-specific settings in configs take precedence over default_config, which takes precedence over system defaults.Your answerB.List the two tool names in an allowed_tools array on that server's mcp_servers entry.tool_configuration.allowed_tools belongs to the deprecated mcp-client-2025-04-04 version. Under mcp-client-2025-11-20, tool selection lives in the tools array.Your answerC.Set enabled to false on the nine unwanted tools inside the mcp_servers entry.Per-tool configuration lives on the mcp_toolset entry, not on the mcp_servers entry, which carries only type, url, name and authorization_token.Your answerD.Declare two separate mcp_toolset entries, one for each tool that stays reachable.Each MCP server can be referenced by exactly one toolset. A second toolset naming the same server fails validation.

   Written from platform.claude.com/docs/en/agents-and-tools/mcp-connector

2. You connect a remote Jira MCP server through the Messages API MCP connector (anthropic-beta: mcp-client-2025-11-20), declaring the server in mcp_servers and a matching mcp_toolset entry in tools. You write a dispatcher that scans each response for tool_use blocks and sends back tool_result blocks for the Jira calls. In testing, responses clearly show Jira tools being invoked and answered, yet your dispatcher never fires. What is the most likely cause?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: C is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.The connector emits tool_use blocks only when the request also sets tool_choice to {"type": "any"}.tool_choice controls whether Claude must emit a call; it never changes the block type used to record one. MCP calls use mcp_tool_use regardless of the choice setting.Your answerB.The Jira calls surface as server_tool_use blocks, the shape Anthropic-executed tools such as web search use.server_tool_use is the block for Anthropic's own server tools (web search, code execution, tool search). The MCP connector defines two distinct types, mcp_tool_use and mcp_tool_result, documented as 'two new content block types'.The answerC.The calls arrive as mcp_tool_use blocks whose mcp_tool_result blocks are already present, run server-side.When Claude uses MCP tools through the connector, the response contains an mcp_tool_use block (with the tool name, server_name, and input) followed by an mcp_tool_result block carrying is_error and the result content.Your answerD.The dispatcher must return results keyed by the entry's server_name field instead of by tool_use_id.server_name appears inside the mcp_tool_use block so a transcript reader can attribute the call. The result is already in the assistant message; the client sends nothing back for it.

   Show the full explanation and doc quotes

   The API executed the call against the remote server during the turn, so there is no tool_use block and nothing for the client to execute or return.

   Written from platform.claude.com/docs/en/agents-and-tools/mcp-connector

3. You integrate a partner's remote MCP server into a Claude Opus 5 (claude-opus-5) application. Your first attempt sends mcp_servers=[{"type": "url", "url": "https://partner.example/mcp", "name": "partner"}] with beta mcp-client-2025-11-20 and is rejected as a validation error. Your organization is also configured for zero data retention and mirrors the workload onto Amazon Bedrock. How should you evaluate this proposal?

   Right. In the game that’s Renown for your Tool Design, MCP and Server Tools building.

   Not that one: C is right. In the game your Tool Design, MCP and Server Tools building would lose a little Renown, and this question would come back later.

   Your answerA.Adding tools=[{"type": "mcp_toolset", "mcp_server_name": "partner"}] fixes the validation error, and the same request then runs unchanged on Bedrock.The mcp_toolset entry is indeed the missing half, but the connector is a hosted feature of the Claude API surface and is not available on Amazon Bedrock. A request that works first-party will not simply run there unchanged.Your answerB.The validation error is caused by the beta header, which must be replaced by managed-agents-2026-04-01; ZDR and Bedrock support follow automatically.managed-agents-2026-04-01 belongs to the Managed Agents endpoints, not the MCP connector, whose beta is mcp-client-2025-11-20. The rejection comes from declaring a server with no corresponding toolset entry, and no beta header changes data-retention eligibility.The answerC.The request needs a matching mcp_toolset entry in tools, but the connector is not ZDR eligible and is unavailable on Bedrock, so a different design is needed.The MCP connector needs both halves of the declaration: mcp_servers for the server URL and a matching tools entry {"type": "mcp_toolset", "mcp_server_name": ...} whose name equals the server name, with beta mcp-client-2025-11-20. mcp_servers alone is a validation error. The connector is available on the Claude API, Claude Platform on AWS and Foundry, is not offered on Bedrock or Vertex, and is not ZDR eligible.Your answerD.The connector requires the team to run its own MCP client process and forward tools as custom definitions, which is what makes it both ZDR eligible and portable to Bedrock.The point of the connector is that Anthropic talks to the remote MCP server for you, with no MCP client of your own. Running your own client and exposing the results as custom tools is a legitimate fallback for ZDR or Bedrock, but it is a different architecture, not the connector's requirement.

   Written from platform.claude.com/docs/en/agents-and-tools/mcp-connector