MCP tools get per-server description limits, timing marks and server ids

What

MCP (Model Context Protocol) servers supply extra tools to Claude Code. Several parts of how these tools are listed and shown changed.

• Tool listing now works out a separate maxDescriptionLength and maxLoadedDescriptionLength for each server, instead of one global limit, and passes them into each tool. The count of over-long descriptions is now measured against the per-server limit.
• Tool listing also adds a serverIdStamp.
• The display data for each tool call now carries server_id (the claude.ai connector id), server_id_stamp (an X-MCP-Server-ID stamp for dynamic http entries) and the bare tool_name. Before, it held only display_name, server_display_name and icon_url.
• MCP tool calls accept a new timing option and mark 'approval', 'own' and 'external' phases. WebFetch and ToolSearch mark timing segments too. This measures where time is spent inside Claude Code.

The stock WebFetch description captured under this release is unchanged.

Why

How much of a tool's description is cut off can now differ from one server to another, which changes what Claude sees. Programs that show tool calls can use the new ids to display a call as belonging to its connector.