Building the MCP Search Tool for any model

Warp's AI agent has been optimized to use the MCP standard for interacting with external tools more efficiently. Previously, preloading all available tools bloated the context window by as much as 50,000 tokens, even though most were unused for a given task. To solve this, a model-agnostic MCP search subagent was developed to dynamically search for and select only the most relevant tools. This approach reduced token usage by 26% in tasks that use MCP and 10% overall, without degrading the agent's performance.