OpenUI

You scaffold a project with a single pnpx command and add OpenUI to the agent code you already have, with no rewrite: it plugs into the Vercel AI SDK, LangChain, LangGraph, Convex, Cloudflare, CrewAI, n8n and Haystack. Your model returns a spec in OpenUI Lang rather than free text, and the client renderer turns that spec into real components from whichever design system you point it at, whether that is the OpenUI design system, shadcn/ui, Material UI, Ant Design, Radix, Chakra, Mantine or daisyUI. The same spec renders across web frameworks, so one implementation covers React, Next.js, Vue, Svelte, Angular, Nuxt, Remix and Astro. For production you can route calls through Gateway, which keeps your existing Chat Completions or Responses SDK, repairs invalid model output as it streams before a user sees it, fails over to another provider during an outage while holding the same model, and reports where usage and spend go. Gateway works on your own provider keys, with no token markup.

• •One Spec, Many Frameworks - A single response format rendered in React, Vue, Svelte, Angular and Lynx, so a UI contract is written once rather than per client
• •Model and Stack Agnostic - Works with OpenAI, Anthropic, Gemini, DeepSeek, xAI, Mistral, Qwen and Ollama, hosted or local, and drops into existing agent frameworks without a rewrite
• •Your Own Design System - Renders through shadcn/ui, Material UI, Ant Design, Radix, Chakra, Mantine, daisyUI or your in-house components rather than a fixed look
• •Output Repair - Gateway catches invalid model output and fixes it mid-stream before it reaches a user, with every repair call included in the plan
• •Provider Failover - Gateway switches providers during an outage while keeping the same model, using your existing credentials and commitments
• •Published Overage Pricing - Per-call overage is on the pricing page at $0.002 on Build and $0.001 on Grow, rather than hidden behind a sales call
• •MIT Licensed Core - The framework is free forever, self-hostable anywhere, with no usage limits from the vendor

Teams building AI products where the answer is something a user has to act on: booking, filtering, comparing, filling in a form, reading a chart. It is the clearest fit for developers already shipping an agent in React or Next.js who want its replies to become interface without hand-writing a renderer per response type, and for anyone who needs the output to match a design system they already maintain. Gateway is aimed at production traffic rather than prototypes, since the free tier's 3,000 calls a month is closer to a demo allowance than a working one, which makes $59 a month the real entry price for a live app. It is not for you if you need a vendor-neutral standard with multiple independent implementations, because that does not exist yet, and it is not for teams whose product is a plain chat transcript, where generative UI adds a rendering layer that earns nothing.