Skip to main content
Metabind is interactive UI for AI agents. Your agent answers with UI components, such as charts, tables, forms, or your own components, instead of text. The answers render in Claude, ChatGPT, and every MCP host, and natively in your own iOS, Android, and web apps. You define the tools; Metabind generates and hosts the MCP server they run on. The agent picks from components you published and fills in their typed fields. It never writes HTML or code. Metabind validates every response against the component’s schema before it renders, keeps your API keys on the server, and runs code isolated on the server and sandboxed on the device.

​
Three ways in

Whichever way a call comes in, it reaches the same generated MCP server and meets the same checks.

​
Your UI components as tools

A Metabind Interactive Tool exposes a UI component to the agent as an MCP tool, with a JSON schema and a component allowlist. When the agent calls the tool, Metabind validates the response against the schema and renders only components the tool allows: product cards, comparison tables, galleries, 3D viewers, or whatever your components define. A response that doesn’t match returns a typed error to the agent instead of rendering.
Rendered Interactive Tool output: a product detail card, a purchasable item card with Buy Now, a 3D model viewer, and a side-by-side product comparison, all returned by agent tool calls and rendered as interactive UI.
Components are written once in BindJS, the open component language for agent UI. In your own apps, they render as SwiftUI on iOS, Jetpack Compose on Android, and React on the web. In Claude, ChatGPT, and other MCP hosts, they render inside the host’s sandboxed frame.

​
Your APIs as tools

A Metabind Data Tool is a typed function the agent can call. You declare an input schema, an output schema, and a handler. The handler runs in a V8 isolate with your secrets injected at runtime, so the model never sees your keys and there’s no server to host or scale. Point a handler at any REST or GraphQL endpoint and it’s an MCP tool. Every MCP App also carries server instructions that tell the agent when to call which tool, tool descriptions on each tool and field, and governance: schema validation, component allowlists, package versioning, and rollback, built in as platform features rather than custom infrastructure.

​
How it works

1

Build your tools

Build Interactive Tools and Data Tools in Metabind Studio’s visual builder and code editor, with live previews on a real device as you go. Or have your AI coding agent write them and push them with the Metabind CLI.
Metabind Studio, the visual development surface for components, tools, and content.
2

Test against the draft endpoint

Every project has two MCP server URLs: a draft endpoint that serves everything (published, modified, and draft) and a production endpoint that serves only published work. Test your tools live in Studio’s Playground, or add the draft URL to Claude while you iterate.
3

Publish to every MCP host

Publishing promotes the draft to mcp.metabind.ai/<your-org>/projects/<your-project>. Claude, ChatGPT, VS Code, and any MCP-compatible host connect with the URL.
4

Run your agent in your own app

The Assistant SDK puts the same tools inside your iOS, Android, or web product, rendered as SwiftUI, Jetpack Compose, or React. The Agent Proxy runs your agent with your system prompt, on Anthropic, OpenAI, or Gemini, with your own model key held server-side.

​
Two distribution surfaces, one MCP App

A single MCP App definition reaches every MCP host (Claude, ChatGPT, others) and embeds inside your own iOS, Android, or web app via the Assistant SDK.
You don’t choose between these. One publish reaches both.

​
What you don’t build

​
How is this different from building your own MCP server?

Building an MCP server from scratch means writing transport and host conformance, schema validation in both directions, secret management, hosting and scaling, and a UI renderer for every surface you want to support, then maintaining all of it as the protocol evolves. Metabind ships that layer as a managed platform. You write the tools that are unique to your product; the protocol, hosting, validation, governance, and native rendering underneath are provided and kept current.

​
Built on open protocols

Metabind aligns with the open Model Context Protocol and its extensions framework (SEP-2133, first proposed as SEP-1724), and Interactive Tools follow SEP-1865 (MCP Apps: Interactive User Interfaces for MCP). Anything you build on Metabind speaks the MCP wire format and runs in any MCP-compatible host. BindJS, its runtime, all three renderers, and the Metabind SDKs are open source under Apache 2.0.

​
Common questions

Yes. Any agent that can connect to an MCP server works, on any model or framework. Add your project’s MCP endpoint beside the servers your agent already uses. It keeps its own model, keys, and hosting. See Custom MCP hosts.
No. Keep it beside Metabind’s: most agents take a list of MCP servers.
Yes. Write BindJS components in Metabind Studio, or have your AI coding agent convert your SwiftUI, Jetpack Compose, or React components to BindJS and push them with the Metabind CLI. Each tool’s component allowlist decides which of your components can render: a product_search tool can render ProductCard but not AdminTable.
Yes, that’s what Data Tools are. Point a Data Tool handler at any REST or GraphQL endpoint, declare the input and output schema, and the platform handles secrets, sandboxing, and schema validation.
The Free plan includes three MCP Apps, 10K tool invocations, Metabind Studio, BindJS, and the Assistant SDK. No credit card. See Pricing.
No. The web renderer (@metabindai/bindjs-react) draws your components inside every MCP host’s sandboxed iframe by default. Native SwiftUI and Jetpack Compose rendering comes with the Assistant SDK, when you run your agent inside your own iOS or Android app.

​
Start building

Try it in Claude or ChatGPT

Add the public Metabind connector, with no account, and see your answers become UI.

Your first MCP App

Five minutes, no code: Oak & Ivory product cards rendering inline in Claude.

​
How the rest of these docs are organized

Your first MCP App

Publish a working MCP App to Claude in five minutes using the Oak & Ivory retail sample.

Core Concepts

The mental model: Types, Components, native rendering, and governance.

Building Your MCP App

The day-to-day authoring workflow: Interactive Tools, Data Tools, BindJS, allowlists, sandbox.

Publishing

Package versioning, publishing the server, rolling back.

Connecting to MCP hosts

Claude, ChatGPT, and any other MCP host.

Assistant SDK

Run your agent inside your iOS, Android, or web app.

Operations

Roles, project visibility, rate limits, SSO.

Sample Apps

Reference MCP Apps you can clone, study, and run end to end.