mcp-gateway-patterns

The mcp-gateway-patterns skill provides a robust architectural framework for designing and integrating Model Context Protocol (MCP) servers within agentic workflows. It addresses the challenge of managing heavy MCP servers that often consume excessive context tokens. By implementing specific isolation strategies—Daemon (SSE), Agent Gateway, and Subprocess—developers can balance startup latency, token overhead, and state sharing requirements effectively. This skill is intended for AI engineers and developers building complex agentic systems who need to decide whether a tool should run as a persistent background service, a lightweight gateway, or an isolated on-demand process. It provides rigorous decision matrices, registration best practices for Claude Code, and clear mappings between user intent and MCP actions. Users will learn how to choose the right pattern based on empirical data, such as token savings thresholds (e.g., tools × 350 > 5,000) and environmental constraints like air-gapped or serverless execution. The skill covers essential topics including MCP registration methods, tool naming conventions, protocol layering (Intent vs. Action), and troubleshooting common issues like visibility and subagent integration. Whether you are building background daemons for persistent state or short-lived subprocesses for low-frequency tasks, this pattern library ensures your MCP integrations remain performant and scalable. Key capabilities include architectural decision guidance, standardizing tool naming formats for plugin and user MCPs, and providing concrete templates for gateway configuration. Practical applications include managing external API integrations, hardware interfaces like SAP2000 OAPI, and high-compute analytical tasks. By following these patterns, you minimize the risk of context bloat, improve agent responsiveness, and ensure your MCP server lifecycle is managed according to best practices.