Vercel launches UI SDK to help AI agents build components
The new open-source library provides a set of behaviors that allow AI coding agents to assemble complex interfaces across multiple web frameworks.
The short version
- Includes over 50 UI behaviors such as keyboard navigation, focus, selection, and drag and drop.
- Supports five major frameworks: React, Vue, Svelte, Solid, and standard Web Components.
- Vercel reported that an agent used 69% fewer tool calls on a settings task using this bridge.
- The library is released under the Apache-2.0 license and is currently in beta at version 0.1.0-beta.1.
Why it matters
This shifts how developers build for AI by providing a standardized 'language' of behaviors that agents can understand. It makes it easier for AI to generate functional, accessible code that works across different web technologies.
A library of behaviors
Vercel Labs recently introduced the UI SDK, an open-source beta library designed specifically for coding agents. Unlike many libraries that provide finished components, this is a headless library that ships behaviors. This means it provides the logic behind an interface rather than the final visual design.
Agents can use these behaviors to assemble accessible components for several popular frameworks. The SDK includes support for React, Vue, Svelte, Solid, and standard Web Components. By providing the underlying logic, the library allows for more flexible creation of user interfaces.
The library includes over 50 distinct behaviors. These range from basic keyboard navigation and focus management to more complex interactions like drag and drop, resizing, and virtualizing lists. It also handles necessary ARIA attributes to ensure the generated components remain accessible to all users.
How the SDK functions
The architecture of the UI SDK relies on primitives defined with a specific function called definePrimitive. A primitive declares the state, an API, and slots for a component. Slots are then used to attach specific attributes, events, and behaviors to different parts of the DOM.
Each framework has its own thin adapter to render these primitives. Because the parts remain the same across all adapters, a single behavior definition can serve multiple different apps. For example, a developer could use the same logic to power a component in both a React and a Svelte project.
The project demonstrates this concept using a Select component. Instead of giving an agent a finished Select, the system provides a popup, a selection model, and a typeahead. These individual building blocks can be combined to create an autocomplete, a combobox, or entirely new components that have not been designed yet.
Agent integration and skills
Vercel designed the SDK so that AI agents can learn and use it quickly. Developers can set up the environment by pasting a prompt into their agent and adding a note to a specific file. This note points the agent toward the UI SDK skills and the documentation provided inside the package.
The system includes a skill that can be installed using a command line tool. This installation provides the agent with the necessary context to understand the library. For projects that do not require a build step, the SDK allows for loading styles and logic directly from a web URL.
The SDK also includes a bridge called createAgent() that runs before a page renders. This bridge allows the agent to see a snapshot of the components, their current states, and the actions available to them. The agent can then execute actions, such as changing a value, without needing to see a screenshot of the screen.
Performance and benchmarks
To test the effectiveness of this approach, Vercel compared their agent bridge against standard browser tools. They used a project settings task where an agent had to rename a project, update members, and change notification settings. Both tests used the Sonnet model through Claude Code.
The results showed that the agent bridge was significantly more efficient. The task time fell by 61%, dropping from 41 seconds to 16 seconds. Additionally, the agent used 69% fewer tool calls than the browser-tools method during the test.
Vercel noted that these figures are directional and based on a single task with three runs. Because the APIs are still in beta, they may change in the future. However, the results suggest that providing agents with structured behaviors is more efficient than letting them interact with the browser directly.
Compatibility and availability
The UI SDK is currently available on npm as six different packages under the Apache-2.0 license. It was published on November 12, 2026, at version 0.1.0-beta.1. To use it, developers can install the core package along with the specific adapter for their chosen framework.
The library has strict peer requirements for the frameworks it supports. It requires React 19+, Vue 3.5+, Svelte 5.33+, and Solid 2.0.0-rc.14+. These requirements ensure that the underlying framework features are modern enough to support the SDK's behavior logic.
While the SDK focuses on behavior, it also provides an optional CSS file. This file gives the components a default look if the developer does not want to provide their own styling. This makes it easier for agents to produce functional and visually coherent interfaces quickly.
What to watch
- How other framework-specific libraries will react to this behavior-first approach.
- Whether more developers adopt the 'agent bridge' over traditional browser-based tool use.
Try it
Check out the ui-sdk.dev website to see examples of how behaviors like 'createPopup' or 'createSelection' work.
Sources and links
- ReportingMarkTechPost: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by MarkTechPost · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
