evantahler

The big idea: you write one action class with a Zod schema, and the framework serves it as an HTTP endpoint, a WebSocket handler, a CLI command, a background task, and an MCP tool with the same validation, same middleware, same `run()` method. The only thing that changes is how the request arrives.

The MCP part is what pushed me to build this. I work on AI agent infrastructure at Arcade (we do authorization and tooling for AI agents), and I kept seeing folks bolt MCP servers onto existing APIs as a separate layer with duplicated auth, duplicated schemas, and duplicated error handling. Keryx makes your API an MCP server by default — OAuth 2.1 + PKCE built in, per-session isolation, and typed errors that agents can actually distinguish.

The stack is opinionated: Bun, Zod, Drizzle ORM, Redis (for Resque-based background tasks and PubSub), PostgreSQL. No compilation step, sub-second startup.

You can try it right now:

  bunx keryx new my-app
  cd my-app
  cp .env.example .env
  bun install
  bun dev

Some things I’d love feedback on:

- Is the “one action, every transport” abstraction too magical, or does it feel natural once you use it? - The MCP-native approach — is this how you’d want to add agent support to a new project? - Anything missing from the comparison table on the homepage?