Agent-Native + Claude Code: Finally! Apps Your AI Agent Can Actually Work In (Open Source)

Agent-Native: https://www.agent-native.com/ GitHub: https://github.com/BuilderIO/agent-native Slides docs: https://www.agent-native.com/docs/template-slides/ In this video, I'll walk through Agent-Native, an open-source TypeScript framework from Builder for building apps where you and an AI agent work in the same interface. I'll start with their Slides app: picking a template, editing it by hand, generating a six-slide deck, and then asking the agent to continue from my manual edits, including a targeted Revise with AI edit, Present mode and export. After that, I'll set up my own copy with the create command, look at how actions and context fit into the code, and use Claude Code to add a review checklist that both the interface and the embedded agent can use. Finally, I'll cover the limitations and where I think this approach is actually useful. -- Key Takeaways: 🧩 Agent-Native is an open-source TypeScript framework for building agents with interactive interfaces: React, your own model, your own repo, and a supported host of your choice. 🖼️ The Slides app lets you start from a prompt or a template like Launch Keynote, and edit headings and slide order directly in the editor. 🤖 After you edit a slide by hand, you can ask the agent to continue, and it works from the deck as it exists now, including your changes. 🎯 Revise with AI targets a selected sentence, so the agent knows exactly which text you mean. 📤 Decks can be presented and exported as HTML, PDF, PPTX or to Google Slides. 💻 You can run your own copy locally with Node 22.22+ and pnpm, then connect Builder, your own provider, or a local Ollama model. 🔀 Actions are defined once with defineAction and can be called from React hooks, the agent, HTTP and MCP. 🧠 Context awareness passes the current view and selection to the agent, so requests stay tied to the object you have open. 🛠️ Because you have the source, a coding agent like Claude Code or Codex can customize the app itself, for example adding a review checklist the UI and the agent share. ⚠️ The example apps are labeled alpha, so you still need to evaluate the model, prompts and actions — but it's a useful direction if you want an agent people can actually work in.