OpenAPI and HTTP Request Tool (I built) is usually enough – no need for MCP (terrydjony.substack.com)

A developer has released openapi-http-tool, an npm package and Vercel AI SDK–compatible utility that lets LLMs call real APIs using a full OpenAPI spec and a simple HTTP fetch tool—essentially the same idea as “GPT Actions.” The tool ingests the OpenAPI YAML/JSON in the system prompt, uses the conversation context to choose which endpoint to call and fill parameters, and provides a system-prompt generator plus a View component for Vercel Agents. The code and examples are on GitHub, and the implementation demonstrates that an LLM + OpenAPI spec + HTTP fetch is sufficient to orchestrate API calls without extra middleware. This matters because it challenges the push for heavyweight MCP (model control plane) stacks just to enable API calling: modern reasoning-capable LLMs can decide which endpoints and parameters to use, so a small, standard fetch tool dramatically lowers integration friction, latency, and infrastructure overhead. If adopted widely, the pattern could encourage a web convention (e.g., /openapi.yml like sitemap.xml) so agents can discover available APIs automatically. Key technical implications include the need to supply the full OpenAPI spec in context, handle authentication and safety at the HTTP/tool layer, and consider standardization and security trade-offs as this lightweight approach scales.