Show HN: Improve Your MCP Servers (agnost.ai)

A Show HN post argues that as MCP (Model Context Protocol) becomes the de facto way LLMs connect to external systems (used by Claude, soon ChatGPT, etc.), most MCP servers are built like REST APIs instead of being optimized for models. The author walks through practical design principles using a GitHub MCP server as an example and even offers a product (Agnost AI) and consulting to monitor and improve MCP deployments. The significance: poorly designed MCPs force models into unnecessary multi-step reasoning, waste tokens, increase errors, and raise latency/costs—so rethinking server design directly improves model reliability and cost-efficiency. Key technical takeaways: design around workflows not endpoint mirrors (e.g., get_repository, create_issue, list_recent_commits), return IDs and any fields the model will need for follow-ups, and write explicit docstrings/tool descriptions so models pick the right tool and avoid retries. Mix actionable tools (fetch/create) with reasoning prompts (analyze a diff). Optimize payloads for token budgets (send commit metadata instead of full diffs) and standardize compact enums/error messages. Instrument everything—tool call frequency, failures, response times, and token usage—to find bottlenecks and fix misuse patterns. These practices enable smoother multi-turn reasoning, lower costs, and more predictable LLM behavior when integrating external systems.