What is UI SDK?
UI SDK, from Vercel Labs, is a TypeScript library of interaction behaviours: keyboard navigation, focus, selection, dismissal, popups, drag and drop, and the ARIA attributes each one needs. They live in @ui-sdk/core, and a component, called a primitive, is defined once and rendered through adapters for React (19 or newer), Vue, Svelte, Solid and Web Components. Ready-made components include Select, Combobox, Field, Tabs, Menu and Date Picker. The project expects an agent to write most of the primitive code, and ships an agent skill and per-version Markdown docs for that.
The second half is an agent bridge. createAgent registers every UI SDK component on the page; snapshot lists each one with its accessible label, value and available actions, act calls those actions, and query searches long lists. The same tools can be handed to your own model or registered with a browser's built-in agent through WebMCP. Because the bridge finds controls by their accessible name, a control a screen reader can identify is also one an agent can find.
The repository was created on 26 October 2026, is Apache 2.0-licensed and has about 6,222 GitHub stars and 217 forks. It is in beta under Vercel's public beta terms; packages were at 0.1.0-beta.13 on 19 November 2026.
Is UI SDK worth it?
The useful idea here is that an agent should not have to read a page the way a browser automation script does. With UI SDK, each component reports its label, current value and the actions it supports, so an agent can set a field labelled "Project name" to "Atlas" in one call instead of finding the box in the accessibility tree, clicking it and typing. The label it matches on is the control's accessible name, so teams that skip aria-label or a proper field label lose both screen-reader users and agents.
Vercel's own benchmark puts numbers on the difference. On a project settings screen built in React, an agent using the bridge needed 7 tool calls, 69.5k tokens, 31 seconds and $0.044, against 17 calls, 291.7k tokens, 53 seconds and $0.139 for the best rival, Base UI driven through browser tools. Vue and Svelte showed cuts of 54% to 55% in cost. These are medians of three runs, with Sonnet in Claude Code, on one task Vercel designed; the benchmark is in the repository, so you can rerun it.
The security design is thoughtful for a beta. Change callbacks receive source "agent" and can cancel a change, an allow callback can refuse any step, controls marked data-agent-confirm ask a person first, and the WebMCP tools carry hints that act is consequential and that results contain untrusted page text. That last point matters: a snapshot can include anything a user typed, which is an opening for prompt injection, so the confirm hooks are worth using on anything destructive.
The catch: everything is new. The repository is under four weeks old, the APIs may change before general availability, and the bridge only sees UI SDK components, so a page that mixes libraries gives the agent a partial view. WebMCP needs a browser that exposes document.modelContext; elsewhere registerTools does nothing. Adopting it means rebuilding existing controls as primitives.
Our take: for a team starting a new React or Vue app that expects agents to operate it, UI SDK is worth piloting on one screen now. For a mature design system built on Radix or shadcn/ui, the sensible move is to watch the beta and fix your accessible names in the meantime, which pays off either way.
Where it earns its place
A front-end lead at a B2B SaaS company adding an in-app assistant
Let the assistant fill in settings forms for the user
Build the settings screen with UI SDK's Field, Select and Combobox, create the bridge at startup and pass agentTools to your model. Use an allow callback and data-agent-confirm on save and delete, so a person approves anything that sticks.
A design-system engineer supporting React and Vue teams
Ship one accessible date picker instead of two diverging ones
Define the primitive once in @ui-sdk/core and render it through each adapter; the testing kit checks keyboard behaviour, focus and ARIA. Budget for API churn while the packages are at 0.1.0 beta.
A QA engineer whose browser agents burn tokens on long forms
Cut the cost of agent-driven end-to-end checks
On screens built with UI SDK, call snapshot and act instead of clicking through the accessibility tree. Vercel's React test fell from 291.7k to 69.5k tokens per run; rerun benchmarks/agent on your own screen first.
Before you commit
- Setup
- npm install @ui-sdk/core and one adapter, such as @ui-sdk/react (React 19 or newer), then add the agent skill with npx skills@latest add vercel-labs/ui --skill ui-sdk. The docs offer a single prompt that does both.
- Integrations
- React, Vue, Svelte, Solid and Web Components; agent tools with JSON Schemas for any model SDK; WebMCP registration through document.modelContext; a server-side route for models that run on your backend.
- Data and privacy
- A client-side library; nothing goes to Vercel. Whatever your agent reads, including text users typed, goes to your model provider, and the tools mark it as untrusted.
- Cost at scale
- No licence cost. Agent runs cost model tokens; in Vercel's benchmark a settings task cost $0.044 to $0.056 per run with Sonnet.
- Limits
- Beta, subject to Vercel's public beta terms. Snapshots list the first 100 components and queries return up to 50 choices; query never opens a closed list.
What’s good
- One component definition renders in React, Vue, Svelte, Solid and Web Components
- Agent bridge cut tool calls from 17 to 7 and cost from $0.139 to $0.044 in Vercel's React test
- Allow callbacks, cancellable changes and confirm markers for agent actions
- Apache 2.0, with docs and an agent skill shipped per version
What’s not
- Beta under four weeks old; APIs may change
- Benchmark is Vercel's own single task
- Bridge sees only UI SDK components, not the rest of the page
- WebMCP works only in browsers that expose document.modelContext
Best for
Front-end teams starting a new app that agents will build and operate, who want accessibility and agent control from one set of components.
UI SDK alternatives
UI SDK: common questions
- Is UI SDK worth it?
- UI SDK makes accessibility the interface an agent drives, a sound idea with a strong vendor benchmark behind it, in a four-week-old beta. We rate it 3.9 out of 5.
- How much does UI SDK cost?
- Free and open source under Apache 2.0, with no paid tier. Agent use costs whatever your model provider charges.
- What are the downsides of UI SDK?
- Beta under four weeks old; APIs may change; benchmark is Vercel's own single task; bridge sees only UI SDK components, not the rest of the page; webMCP works only in browsers that expose document.modelContext.
- Who is UI SDK best for?
- Front-end teams starting a new app that agents will build and operate, who want accessibility and agent control from one set of components.
- How much does UI SDK cost for a team or in production?
- No licence cost. Agent runs cost model tokens; in Vercel's benchmark a settings task cost $0.044 to $0.056 per run with Sonnet.
- Is UI SDK safe for company data?
- A client-side library; nothing goes to Vercel. Whatever your agent reads, including text users typed, goes to your model provider, and the tools mark it as untrusted.
- What are the best UI SDK alternatives?
- The closest alternatives we have reviewed are shadcn/ui, Browser Use and Stagehand.
Answers checked Nov 20, 2026.