Tool Search
A context optimization mechanism in which an AI coding assistant loads only tool names at startup and retrieves and loads the detailed tool specifications on demand, reducing the token overhead of tool definitions.
Timeline (last 90 days)
According to the video demo, when the number of tools reaches hundreds, tool search is enabled by default and Claude loads only the subset of tools needed for the current task.
When Claude Code starts, it only loads the names of MCP tools, and only searches for and loads the full documentation of the corresponding tool when a task needs it; this mechanism is called 'tool search'.
According to Anthropic's estimates, a typical five-server MCP configuration spends about 55,000 tokens on tool documentation, and tool search can cut this overhead by more than 85%.
In Claude Code, once a tool's documentation has been loaded into the current session, it remains until the session ends, causing context consumption to accumulate gradually over long sessions (session stickiness).
The savings from tool search are most pronounced in short sessions, while in long sessions with many tools they are gradually eroded by session stickiness.
If Tool Search's retrieval recall is inaccurate, the model may not get the correct tool, and the semantic clarity of tool names and namespace descriptions directly affects the final result.
All Facts (6)
According to the video demo, when the number of tools reaches hundreds, tool search is enabled by default and Claude loads only the subset of tools needed for the current task.
50%UnverifiedWhen Claude Code starts, it only loads the names of MCP tools, and only searches for and loads the full documentation of the corresponding tool when a task needs it; this mechanism is called 'tool search'.
50%UnverifiedAccording to Anthropic's estimates, a typical five-server MCP configuration spends about 55,000 tokens on tool documentation, and tool search can cut this overhead by more than 85%.
50%UnverifiedIn Claude Code, once a tool's documentation has been loaded into the current session, it remains until the session ends, causing context consumption to accumulate gradually over long sessions (session stickiness).
50%UnverifiedThe savings from tool search are most pronounced in short sessions, while in long sessions with many tools they are gradually eroded by session stickiness.
50%UnverifiedIf Tool Search's retrieval recall is inaccurate, the model may not get the correct tool, and the semantic clarity of tool names and namespace descriptions directly affects the final result.
50%