BuilderIO / agent-native

查看原文 / View Original 编辑 / Edit

BuilderIO / agent-native

Agent-Native Open-source framework for agentic applications you own. Agent-Native 是一个用于构建你所拥有的智能体应用（Agentic Applications）的开源框架。

Don’t choose between rich user interfaces and autonomous agents. Every Agent-Native app is both. 无需在丰富的用户界面和自主智能体之间做选择。每一个 Agent-Native 应用两者兼备。

Agents and UIs — Fully Connected 智能体与 UI — 全面互联

The agent and the UI are equal citizens of the same system. Every action works both ways — click it or ask for it. 智能体和 UI 是同一系统中的平等成员。每一个操作都支持双向交互——你可以点击它，也可以通过对话请求它。

Everything syncs — Agent and UI share one database and one state. Changes from either side show up instantly on the other. 一切皆同步——智能体和 UI 共享同一个数据库和状态。任何一方的更改都会立即在另一方显示。

Real-time multiplayer — Humans and agents collaborate in the same document simultaneously: CRDT merging, live presence (cursors, selection rings, who’s on which slide), and the agent as a first-class peer editor. 实时多人协作——人类和智能体可以在同一文档中同时协作：支持 CRDT 合并、实时在线状态（光标、选择框、谁在查看哪一页），并将智能体视为一等公民的对等编辑器。

Works on any SQL database and any host, including serverless. 适用于任何 SQL 数据库和任何托管环境，包括 Serverless。

Context-aware — The agent knows what you’re looking at. Select text, hit Cmd+I, and tell it what to do. 上下文感知——智能体知道你正在查看的内容。选中文字，按下 Cmd+I，然后告诉它要做什么。

Per-user workspace — Skills, memory, instructions, sub-agents, and MCP servers — SQL-backed, customizable per user. 用户专属工作区——技能、记忆、指令、子智能体和 MCP 服务器——基于 SQL 后端，可按用户进行自定义。

Claude-Code-level flexibility, SaaS-grade economics. 具备 Claude-Code 级别的灵活性，以及 SaaS 级别的经济性。

Agents call agents — Tag another agent from any app. They discover each other over A2A and take action across your stack. 智能体调用智能体——在任何应用中标记另一个智能体。它们可以通过 A2A（Agent-to-Agent）协议相互发现，并在你的技术栈中执行操作。

Reusable integrations — Connect a provider once in Dispatch, keep secret values in the vault, then grant apps like Brain, Analytics, Mail, and Dispatch access to the shared account metadata and credential refs. 可复用的集成——在 Dispatch 中连接一次提供商，将密钥保存在保险库中，然后授权 Brain、Analytics、Mail 和 Dispatch 等应用访问共享的账户元数据和凭据引用。

Three shapes — Build the same agent as a headless API, a rich chat experience, or a full application where agent and UI stay in sync. 三种形态——构建同一个智能体，既可以作为无头 API（Headless API），也可以作为丰富的聊天体验，或者是一个智能体与 UI 保持同步的完整应用程序。

Apps that improve themselves — Your apps get better on their own. The agent can add features, fix bugs, and refine the UI over time. 自我进化的应用——你的应用会自主进化。智能体可以随着时间的推移添加功能、修复错误并优化 UI。

Any database, any host — Any SQL database Drizzle supports. Any hosting target Nitro supports. No lock-in. 任何数据库，任何主机——支持 Drizzle 支持的任何 SQL 数据库，以及 Nitro 支持的任何托管目标。无厂商锁定。

Bring the agent surface you need — MCP-compatible hosts can call your apps, coding agents can install skills, native chat renders reusable app outputs, and BYO agent runtimes can stream into the Agent-Native chat shell. 按需引入智能体界面——兼容 MCP 的主机可以调用你的应用，编程智能体可以安装技能，原生聊天界面可以渲染可复用的应用输出，自带（BYO）智能体运行时可以流式传输到 Agent-Native 聊天外壳中。

The framework for agent-native apps

Agent-native 应用框架

Agent-Native is an open-source framework for building robust agents that can act inside real apps, not just chat next to them. It gives you primitives for product-grade agentic software: shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together. Agent-Native 是一个开源框架，用于构建能够直接在真实应用内部操作（而不仅仅是在旁边聊天）的健壮智能体。它为你提供了产品级智能软件所需的基础组件：共享操作、基于 SQL 的状态、身份验证、工具、技能、任务、可观测性以及协同工作的 UI 界面。

Backend agnostic: bring your own database, hosting provider, model stack, and app code. 后端无关：自带数据库、托管提供商、模型栈和应用代码。

// One action powers UI, agent, HTTP, MCP, A2A, and CLI.
// 一个操作即可驱动 UI、智能体、HTTP、MCP、A2A 和 CLI。
export default defineAction({
  schema: z.object({
    emailId: z.string(),
    body: z.string(),
  }),
  run: async ({ emailId, body }) => {
    await db.insert(replies).values({ emailId, body });
  },
});

Actions — Define work once. Use it from UI, agent, API, MCP, A2A, and CLI. 操作——定义一次工作，即可在 UI、智能体、API、MCP、A2A 和 CLI 中使用。

Agent runtime — Chat, tools, skills, memory, jobs, observability, and handoffs ship together. 智能体运行时——聊天、工具、技能、记忆、任务、可观测性和交接功能开箱即用。

Backend agnostic — Plug in any Drizzle-supported SQL database and Nitro-compatible host. 后端无关——接入任何 Drizzle 支持的 SQL 数据库和 Nitro 兼容的主机。

One agent, three product shapes

一个智能体，三种产品形态

Agent-Native primitives let you choose how much UI to put around an agent without rebuilding the agent contract: Agent-Native 的基础组件让你无需重构智能体契约，即可选择为智能体配置多少 UI：

ShapeWhat you shipSame primitives underneath
HeadlessCall the agent and actions from code, CLI, HTTP, MCP, or A2A.defineAction, auth, skills, memory, jobs, observability
Rich chatA standalone or embedded chat with native tables, charts, approvals, setup flows, and tool results.Shared chat runtime, BYO runtime adapters, action-declared native renderers
Whole appA full SaaS/product UI where chat can start central, move to the sidebar, and stay synced with app state.SQL state, actions, context awareness, deep links, live sync
形态你交付的内容底层共享的基础组件
无头 (Headless)从代码、CLI、HTTP、MCP 或 A2A 调用智能体和操作。defineAction, auth, skills, memory, jobs, observability
富聊天 (Rich chat)带有原生表格、图表、审批、设置流程和工具结果的独立或嵌入式聊天。共享聊天运行时，自带运行时适配器，操作声明的原生渲染器
完整应用 (Whole app)完整的 SaaS/产品 UI，聊天可以从中心开始，移动到侧边栏，并与应用状态保持同步。SQL 状态，操作，上下文感知，深度链接，实时同步

Protocols come with the framework instead of becoming separate integrations per feature. Today that means A2A, MCP, MCP Apps, standard remote MCP OAuth, MCP clients, HTTP/CLI action calls, native chat widgets, AgentChatRuntime adapters, standard OpenAI, AG-UI, Claude Agent SDK, and Vercel AI SDK chat runtime connectors, and deep links all hang off the same action surface. 协议随框架内置，而不是为每个功能单独集成。目前这意味着 A2A、MCP、MCP Apps、标准远程 MCP OAuth、MCP 客户端、HTTP/CLI 操作调用、原生聊天组件、AgentChatRuntime 适配器、标准 OpenAI、AG-UI、Claude Agent SDK、Vercel AI SDK 聊天运行时连接器以及深度链接，全部挂载在同一个操作界面上。

ACP is best understood as the coding-agent/editor interoperability protocol, not the general BYO app-chat runtime. ACP 最好被理解为编程智能体/编辑器的互操作协议，而不是通用的自带应用聊天运行时。

For the full decision guide — headless, rich chat on the built-in agent, rich chat on your own agent, embedded sidecar, or full app — see Agent Surfaces. 有关完整的决策指南（无头模式、内置智能体的富聊天、自定义智能体的富聊天、嵌入式侧边栏或完整应用），请参阅 Agent Surfaces。

To connect Claude, ChatGPT, Codex, Cursor, OpenCode, GitHub Copilot / VS Code, or another MCP host to your hosted app, see the External Agents guide. 要将 Claude、ChatGPT、Codex、Cursor、OpenCode、GitHub Copilot / VS Code 或其他 MCP 主机连接到你的托管应用，请参阅 External Agents 指南。

Try it with a skill

通过技能尝试

Don’t want to scaffold a whole app yet? Add visual planning and PR recaps to Claude Code, Codex, Cursor, Pi, OpenCode, GitHub Copilot / VS Code, and similar agents with one command: 不想现在就搭建整个应用？只需一条命令，即可为 Claude Code、Codex、Cursor、Pi、OpenCode、GitHub Copilot / VS Code 等类似智能体添加可视化规划和 PR 回顾功能：

npx @agent-native/core@latest skills add visual-plan

You get two slash commands that upgrade how your agent plans and reports its work: 你将获得两个斜杠命令，升级你的智能体规划和报告工作的方式：

  • /visual-plan — before the agent writes code, it opens a structured, reviewable plan document instead of a wall of text: inline diagrams, UI wireframes and prototypes, file-by-file implementation maps, and annotations you can comment on and approve.
    • /visual-plan — 在智能体编写代码之前，它会打开一份结构化、可审查的计划文档，而不是一大段文字：包含内联图表、UI 线框图和原型、逐文件的实现地图，以及你可以评论和批准的注释。
  • /visual-recap — after changes land, it turns a PR or git diff into a high-altitude visual recap: schema, API, and file changes rendered as grounded before/after blocks with a shareable review link, instead of scrolling a raw diff.
    • /visual-recap — 在更改落地后，它会将 PR 或 git diff 转换为高层级的可视化回顾：模式、API 和文件更改被渲染为直观的“修改前/修改后”区块，并附带可分享的审查链接，无需滚动查看原始 diff。

See the Skills Guide for more skills and local installs. 查看 Skills Guide 以获取更多技能和本地安装信息。

Templates

模板

Start with a full featured template. Each one is a complete, 100% free and open-source SaaS app — cloneable, not scaffolded — except you own the code and can customize everything. 从功能齐全的模板开始。每一个都是完整、100% 免费且开源的 SaaS 应用——可克隆，而非简单的脚手架——你拥有代码并可以自定义一切。

  • Calendar: Agent-Native Google Calendar, Calendly. Manage events, sync with Google Calendar, and share a public booking page with AI scheduling.
    • 日历： Agent-Native Google Calendar, Calendly。管理事件，与 Google Calendar 同步，并分享带有 AI 调度功能的公共预订页面。
  • Content: Open-source Obsidian for MDX. Edit local Markdown/MDX files, generate rich interactive custom blocks, and draft, rewrite, or publish with an agent.
    • 内容： 开源的 MDX 版 Obsidian。编辑本地 Markdown/MDX 文件，生成丰富的交互式自定义区块，并使用智能体进行草拟、重写或发布。
  • Plans: Visual plan mode for coding agents. Install /visual-plan and /visual-recap so your coding agent can plan before it builds and recap changes after they land — high-level code reviews with diagrams, wireframes, annotations, and review links.
    • 计划： 编程智能体的可视化计划模式。安装 /visual-plan 和 /visual-recap，让你的编程智能体在构建前进行规划，并在更改落地后进行回顾——提供带有图表、线框图、注释和审查链接的高级代码审查。
  • Slides: Agent-Native Google Slides, Pitch. Generate and edit React-based presentations.
    • 幻灯片： Agent-Native Google Slides, Pitch。生成并编辑基于 React 的演示文稿。
← 返回新闻列表 / Back to news list