What I learned building MCP servers for LinkedIn, Medium, Kaggle and X Over the past year I've open-sourced a series of MCP servers that let AI assistants like Claude work directly with LinkedIn, Medium, Kaggle and X. The biggest lesson had nothing to do with the APIs. It was how the tools are described. Three things made the clearest difference to whether an agent picked the right tool: - Action-first names like search_posts or get_profile, instead of vague ones like posts or query. - Descriptions that say when to use a tool, not just what it does. - Short, precise parameter docs with examples of valid values. With those changes, wrong tool calls dropped noticeably and agents needed far fewer retries. If you're adding AI agents or LLM features to your product and want them to work reliably in production, I'd be glad to help. Code: github.com/Dishant27