Builder’s Agent-Native Beta Makes One Action Serve UI, Agent, API, CLI, MCP and A2A

Builder.io launched Agent-Native in beta on October 14, 2026, as an open-source TypeScript framework for agents that ship with a real application UI instead of a chat pane alone. The framework repository BuilderIO/agent-native is public on GitHub; the @agent-native/core package on the main branch is version 0.209.0 and is licensed MIT, while the repository root package.json declares ISC. Builders can scaffold locally with npx @agent-native/core@latest create or build hosted apps on Builder.io.

The pitch is architectural, not another agent loop wrapper. Agent-Native treats the agent and the UI as equal clients of the same capability layer. You define an action once. The agent calls it as a tool. The UI calls the same implementation from application code. Both paths share validation, permissions and business logic. Builder says that single definition also fans out to an HTTP API, a CLI, an MCP server and the Agent-to-Agent (A2A) protocol—six surfaces from one schema.

Why a UI belongs next to the agent

Builder’s argument starts from ordinary knowledge work. Slide editors, design tools and dashboards exist because people need to see state and nudge a detail without rewriting a prompt. Pure chat agents force every tweak through tokens: change a color, wait, pay, hope nothing else moved. Agent-Native keeps the prompt path for open-ended work and the UI path for precise edits, with both writing the same underlying data.

That pairing is what Builder calls an agentic app: an agent that comes with its own application, not a sidebar bolted onto a human-only product. The UI advertises what the agent can do. Context already lives in the app—current page, selected record, active view—so “fix this” is grounded. Results appear in the interface as the agent works, and a person can keep editing, select a region, or hand the task back.

Three shared foundations, six outbound surfaces

Under the hood, Builder describes three shared foundations. Shared actions are the capability contracts. Shared data keeps agent writes and human edits coherent. Shared application state tells the agent what is on screen so it does not have to guess from a transcript alone.

Compared with computer-use agents that click through pixels, Agent-Native agents call the same actions the UI uses. Builder’s claim is fewer tokens and less brittle screen parsing, plus the ability to run without a display—from Slack, on a schedule, or at another agent’s request. Messaging surfaces named in the post include Slack, Microsoft Teams, Discord, email, Telegram and WhatsApp. Tool connections include Notion, Linear and GitHub through MCP and built-in connectors.

A2A support means one Agent-Native app can delegate a task to another agent that keeps its own tools and context. MCP exposure means Claude, ChatGPT, Cursor or Codex can call the app’s actions directly, and hosts that support MCP Apps can render the app UI inside the conversation. That is the “six surfaces” story in practice: UI, agent, HTTP, CLI, MCP and A2A sharing one action graph.

Team primitives and the toolkit apps

Beta includes shared instructions, skills and memory, plus authentication, permissions, organizations, multi-tenancy and live sync. Sensitive steps can require human approval. History tracks whether a change came from a person, a teammate or the agent. An optional toolkit supplies editing, collaboration, version history and app-shell pieces used by Builder’s own reference apps—Clips, Slides and Design—which are themselves open-source agentic apps built on the framework.

Model choice is deliberately loose. Bring keys for Anthropic, OpenAI, Google Gemini, OpenRouter or any OpenAI-compatible endpoint, run Ollama locally, or use Builder’s managed connection. Switching models does not rewrite actions or UI. Local projects keep code in your repository, attach PostgreSQL, and deploy to Node, Vercel, Netlify, Cloudflare or AWS. Existing React apps can add an agent panel without a greenfield rewrite.

Beta means architecture first, not a finished product suite

Agent-Native is explicitly a beta. The October 14 post is a framework launch, not a migration mandate. The useful test is concrete: pick one workflow your team already knows, define its capabilities as actions, and see whether the same contracts serve both a human UI and an agent without forked business logic.

The news is less “another TypeScript SDK” and more a public bet that the durable unit of agent software is a shared action graph with a face. An MIT-licensed core package, a public monorepo and day-one MCP/A2A exposure make that bet inspectable. Whether the six-surface model reduces token waste in production is empirical; the October 14 beta is when teams can start measuring it on their own apps.