OpenAPI and HTTP Tool Call is usually enough, no MCP is needed (blog.terrydjony.com)

A developer demonstration argues that OpenAPI + a generic HTTP tool call is usually enough for AI agents to interact with APIs—no dedicated “MCP” server required. Using the Vercel AI SDK, they built an OpenAPI tool (validated with a z.object input schema) that extracts baseUrl from the OpenAPI servers section, picks the correct path and HTTP method, fills query parameters/body/headers (including array handling via comma-joins), and performs fetch requests with JSON body handling, content-type detection, and structured progress/error/success yields. A complementary system prompt generator instructs the agent to read the spec, identify required parameters and authentication from components.securitySchemes, validate types, and call the tool with a complete request. The demo (a weather agent using OpenMeteo) shows an agent reasoning over the spec, selecting endpoints, and returning API results to users. This approach matters because it lowers friction for API integration with AI agents: rather than convincing every API provider to run a special proxy, developers can publish a discoverable /openapi.yml and let agents parse and call endpoints directly. That accelerates agent capabilities and adoption while centralizing effort on making specs machine-readable and secure (clear auth schemes, accurate servers, parameter typing). Practical implications include simpler tooling, fewer custom wrappers, and the need to standardize spec discoverability and strong auth/validation practices to avoid misuse.