vis-Neuclea: Scoped Agent Infrastructure for APIs

THE WEB WASN'T BUILT FOR AGENTS Modern web architecture was engineered for humans. Forcing AI agents into it creates three bottlenecks: bloated ReAct loops that burn thousands of tokens per query, credentials that have to live on third-party backend servers, and bespoke API wrappers that slow every new integration. THE SOLUTION: VIS + NEUCLEA VisNeucla turns any OpenAPI 3.x spec into a conversational AI agent in minutes — no boilerplate, no glue code. Vis is a TypeScript CLI that parses an OpenAPI spec through a strict one-way pipeline (core → parser → templates → generator → cli) and emits a typed, production-ready MCP server: Zod-validated tools, automatic parameter routing across query/path/header/body, clean JSON Schema for tool discovery, and an optional containerized observability stack (Prometheus, Loki, Tempo, Grafana) for metrics, logs, and tracing. Neuclea is a Go WebSocket gateway that runs a round-based ReAct agent (max 3 rounds) against that MCP server. Independent tools in a round execute in parallel via goroutines. Every design choice is tuned for tokens: first-sentence-only tool descriptions, required-params-only listings, name-only history summaries, and capped output at each planning and formatting step. A Markov-chain predictor learns tool-call sequences to power autocomplete, and the MCP client pool handles per-endpoint rate limiting, retries with backoff, and cold-start detection. THE RESULT A typical two-round query costs ~1,800–2,200 tokens — roughly 5x less than an unconstrained agent loop — while streaming live reasoning steps over WebSocket for a transparent, responsive console. TECH STACK Go 1.25, TypeScript, gorilla/websocket, Zod, pnpm workspaces, Fireworks AI (MiniMax M2P7 / GLM series), Ollama, Model Context Protocol, OpenAPI 3.x, Prometheus, Grafana, Loki, Tempo, Promtail.