Tool Search

What it does

The Tool Search skill enables AI agents to efficiently discover and load relevant tools from a catalog, scaling beyond traditional limitations. Instead of loading all tools upfront (which consumes significant tokens), this skill uses embeddings for semantic matching to identify only the necessary tools based on the agent's task context. This approach reduces token consumption by up to 85% compared to traditional methods, allowing for larger tool libraries and improved accuracy in tool selection.

When to use it

• When tool definitions are consuming more than 10K tokens.
• When a tool library contains 10 or more tools.
• When experiencing issues with the accuracy of tool selection.
• When building MCP-powered systems that utilize multiple servers (e.g., repo, github).

Key capabilities

• Semantic Tool Search: Finds relevant tools based on task context using embeddings.
• Embedding-Based Matching: Uses embeddings for accurate tool discovery.
• On-Demand Loading: Loads tools only when needed, minimizing token consumption.
• Context Efficiency: Reduces tool definition tokens by 90%+.

Example prompts

• "Create a pull request for my changes." (The agent will search for "github pull request creation")
• "Search customer orders by date range."
• "I need to read a file from disk."

Tips & gotchas

• Configuration Required: Requires MCP configuration with betaFeatures and toolSearch settings enabled.
• Clear Tool Descriptions: Ensure tools have clear names and descriptions for effective semantic matching.
• Always Load Critical Tools: Consider always loading essential tools like read_file, write_file, search_code, create_pull_request, or get_issue to ensure availability when needed.