MCPulse

TypeScript, Node, React, Postgres. Fastify, Astro, Supabase, Vercel. Generalist by necessity — solo means SDK, API, dashboard, billing and docs are all the same job.

How models actually read tool schemas. Next step is a model-in-the-loop study: same 4,951-server corpus, but with a model selecting tools, to see which schema signals predict a real mistake.

MCPulse: an SDK that runs inside your MCP server and reports first-call success, retries, empty results, dead tools and schema cost. Plus a free browser checker for tool schemas.

Talk to me if you maintain an MCP server and can't tell whether models understand your tools. Also happy to argue about methodology — someone already found a real hole in my last study.