Builder.io wants the agent to be a real user of your app
Builder.io just open-sourced agent-native, a framework with a simple but sharp premise: stop bolting an AI chatbot onto your app, and build the app so an agent is a first-class user from the start. It's trending on GitHub this week and it's built directly on Anthropic's Claude Agent SDK.
The core idea is shared actions. You define an action once and it works everywhere, your UI button, an agent calling it, an HTTP endpoint, an MCP tool, a CLI command all hit the same code path. Underneath it gives you SQL-backed state, identity, tools, skills, jobs, and observability in one stack, plus real-time multiplayer between humans and agents over CRDTs. So a person and a Claude Code or Cursor session can edit the same app live, the way two people share a Google Doc.
Why it matters: most AI apps today are a normal app with a chat box stapled on, and the agent can only see whatever the chat window exposes. agent-native flips it, the agent gets the same verbs the UI has, no more and no less. That's the difference between an agent that can suggest and an agent that can actually do.
It's early, about a thousand stars, version 0.63, and it's one opinionated bet among many about what agent-native software means. But the framing is the right one, and coming from the team behind Qwik and Mitosis, it's worth watching. You can spin one up with npx @agent-native/core@latest create.
https://github.com/BuilderIO/agent-native
← Back to all articles
The core idea is shared actions. You define an action once and it works everywhere, your UI button, an agent calling it, an HTTP endpoint, an MCP tool, a CLI command all hit the same code path. Underneath it gives you SQL-backed state, identity, tools, skills, jobs, and observability in one stack, plus real-time multiplayer between humans and agents over CRDTs. So a person and a Claude Code or Cursor session can edit the same app live, the way two people share a Google Doc.
Why it matters: most AI apps today are a normal app with a chat box stapled on, and the agent can only see whatever the chat window exposes. agent-native flips it, the agent gets the same verbs the UI has, no more and no less. That's the difference between an agent that can suggest and an agent that can actually do.
It's early, about a thousand stars, version 0.63, and it's one opinionated bet among many about what agent-native software means. But the framing is the right one, and coming from the team behind Qwik and Mitosis, it's worth watching. You can spin one up with npx @agent-native/core@latest create.
https://github.com/BuilderIO/agent-native
Comments