mcp-context-optimizer_skill

Overview

This skill optimizes MCP tool context loading by using progressive disclosure and hierarchical indexing to prevent context window bloat. It keeps a small category index always loaded, fetches tool indexes on demand, and only loads full tool schemas when required. Use it when working with large MCP collections to reduce token use and speed up tool selection.

How this skill works

The skill maintains a three-level hierarchy: a lightweight category index (always loaded), per-category tool indexes loaded on demand, and full tool schemas fetched only when executing or inspecting a specific tool. Incoming user intent is parsed to match a category, a scoped search runs inside that category, and the full tool schema is retrieved only for the chosen tool. Caching and simple routing rules reduce repeated lookups and lower token costs.

When to use it

• Working with large MCP sets (dozens to hundreds of tools) to avoid context bloat
• When user queries include triggers like "find tool", "search mcp", or "which tool"
• Before running multi-step workflows that touch multiple categories
• When response latency or token cost is a concern
• During agent orchestration to limit per-agent context size

Best practices

• Always start by matching the user intent to a category before broad searching
• Call mcp_search_tools with a category filter before requesting full tool schemas
• Cache frequently used tool schemas within the conversation context
• Bundle related tool calls into a single workflow to reduce repeated lookups
• Avoid loading all tool definitions up front; load schemas only when needed

Example use cases

• User asks to check container logs: match to docker, search for logs tool, fetch schema, then execute docker_logs
• Deploy and verify workflow: locate compose_up, network_port_check, docker_logs, and health_check across categories
• Ambiguous request: run a broad mcp_search_tools query, review results, then narrow to a category
• Agent orchestration: lead agent uses optimizer to hand sub-agents minimal schemas needed for tasks
• Exploration mode: list categories first to conserve tokens and guide follow-up queries