wtf is omnimcp?

OmniMCP was built for the OKX.AI Genesis Hackathon and aims to solve a specific onboarding problem for AI agent marketplaces. Many useful services already exist as ordinary web APIs, described by an OpenAPI or Swagger specification, but AI agents typically need a different kind of interface called MCP to talk to tools automatically. OmniMCP reads an existing API's specification and automatically generates a working MCP style agent surface for it, so that an old web API can be plugged into an agent economy without anyone hand writing integration code. The tool fetches and normalizes OpenAPI 3.x or Swagger 2.0 specifications, then uses an AI model, specifically Gemini 2.5 Flash through OpenRouter, to translate each REST endpoint into a clearly described tool definition that other AI agents can understand and call. Once generated, each agent exposes a JSON-RPC endpoint with methods for listing available tools and calling them, letting one AI agent invoke another agent's underlying API through this generated layer. Behind the scenes, a proxy engine intercepts these calls, rebuilds the correct HTTP request to the original API, and returns the result. The project also wires in monetization and verification pieces tied to the OKX.AI ecosystem: generated tools can be wrapped so each call is billed through a pay per call protocol, and the schema of each agent gets a cryptographic hash computed for verification on a blockchain called X Layer. For the hackathon demo itself, the app includes a fallback mode that swaps in prewritten example responses if the AI service is temporarily overloaded, so the live demonstration keeps working. Running it locally requires Node.js 18 or newer and a free OpenRouter API key. It stores generated agents in local files rather than a database, so no separate setup is needed to try it out. The project is released under the MIT license.