Anvil is a backend framework built for a world where APIs get called by both browsers and AI agents. Write one route file, you get a REST endpoint, an MCP tool, and an A2A skill, all sharing the same schema, checked at build time. No hand-wiring MCP SDKs on top of Express, no duplicating validation across three layers.

Hey Product Hunt 👋 I built Anvil because I wanted an Express-level flexible backend framework with Next.js-style file routing and real compile-time schema validation, without wiring up a dozen separate libraries for auth, rate limiting, observability, and crash recovery on every project. One route file gives you a type-safe REST endpoint out of the box. Auth, sessions, rate limiting, tracing, and crash-resumable checkpointing come with the framework, not bolted on later. Since a lot of what I build these days touches AI, Anvil also has an optional layer for that: expose any route as an MCP tool or A2A skill, or add a full agent runtime with tool-calling, streaming, multi-model fallback, evals, and memory/RAG. None of it is required, it's there when you need it. 200+ tests, still actively hardening a few areas (validation coverage, streaming, vector store adapters, sandboxing), roadmap's open, feedback welcome. Would love for you to try `npx anvil init` and tell me what breaks. 🙏

Finally got to play with Anvil and it's refreshing to have one source of truth for both REST and MCP. One thing I'd love to see is built-in support for streaming responses in the route files, especially for long-running agent tasks where you want to send partial results back without blocking the whole request. That would make it a no-brainer for real-time use cases.

The build-time schema check is genuinely useful, caught a mismatched response type before I even hit run. Love that one route file gives me MCP, REST, and A2A without any glue code.

Anvil SDK

The Next.js of AI backends

The Next.js of AI backends