About this project
Give NitroForge an OpenAPI spec. Get back a verified, working MCP server — typechecked, built, booted, and tested against real protocol calls, in under 15 seconds.Most "AI generates code" tools ask you to trust the model. NitroForge doesn't. We split the pipeline into two fundamentally different kinds of work: judgment and mechanics. The one LLM call in the entire system picks tool names, clusters endpoints, and writes descriptions — things that genuinely require reasoning about what a server should look like. Everything else — the actual input schemas, the field types, the HTTP calls, the validation logic — is derived deterministically from the OpenAPI spec itself and compiled by the real TypeScript compiler. The model literally cannot hallucinate a field name, because it never writes the code that defines one.That split makes NitroForge self-verifying, not self-reported. Every generated server goes through a four-stage machine oracle: real typecheck, real build, real process boot completing a real MCP handshake, and a live replay of every tool against its expected response — diffed, not assumed. If a generated server passes, it's because a compiler and a running process proved it, not because a model said so. Under the hood, NitroForge is itself a full MCP server — tools for the pipeline, Resources for inspecting what's already been built, Prompts for chaining the workflow into one guided action. It's deployed live on NitroCloud, exposes real, callable tools, and takes real traffic.We built this against a framework with contradictions between its own documentation and its actual behavior — decorator conventions, DI resolution order, transport configuration that silently does nothing unless you know the real mechanism. We found those bugs by reading source, not assuming docs, and fixed several before they became silent production failures. NitroForge isn't a wrapper. It's an argument.
Open Innovation track
Solve any real-world problem with AI, regardless of industry or domain.