How
The model chooses tools; the app maps tool results to components. runAgentTUI hosts an agent in a terminal chat.
Vercel AI SDK generative UI (RSC streamUI, AI SDK UI, @ai-sdk/tui)
Model tool calls render React components; plus a canned agent TUI.
ai-sdk.dev ↗repo ↗docs ↗runAgentTUI reference ↗Migrating from RSC to AI SDK UI ↗
- license
- Apache-2.0 · LICENSE file is Apache-2.0 (Copyright Vercel); GitHub API reports NOASSERTION.
- language
- TypeScript
- backing
- Vercel
- stars
- 27,211 (as of 2026-10-09)
- downloads
- ai 29,092,186/week; @ai-sdk/react 9,563,237/week; @ai-sdk/rsc 113,633/week; @ai-sdk/tui 27,329/week (npm, week ending 2026-10-07)
- latest release
- ai@7.0.137; @ai-sdk/rsc 3.0.137; @ai-sdk/tui 1.0.138, 2026-10-09
- first release
- 2023-06
ignore
Code-first generative UI and a fixed chat TUI; json-render is Vercel's data-first answer and the better comparison.
Ignore it for this purpose.
the steelman: the best honest case for it
The AI SDK is where most TypeScript LLM apps start (29M weekly downloads), and its generative-UI story is the simplest one there is: let the model call a tool, render a real component for the result, stream it. No new UI language for the model to learn, no catalog DSL, full React and design-system fidelity, type-safe tool parts in useChat, and with @ai-sdk/tui the same agent runs in a polished terminal chat in three lines. For teams that just need 'show a weather card when the model calls getWeather', it is hard to beat.
scores
| UI as data | 1 scene 5 | Components are React code; only tool args/results are data. |
|---|---|---|
| Agent can drive it | 2 scene 5 | |
| Agent can see it | 1 scene 5 | |
| Small, targeted edits | 1 scene 5 | |
| Time travel | 0 scene 0 | |
| Live data rate | 2 scene 5 | |
| Terminal native | 2 scene 5 | runAgentTUI renders a fixed chat UI in the terminal; no general terminal renderer. |
| Pixel graphics | 3 scene 4 | Whatever React on the web draws. |
| Teaching | 0 scene 3 | |
| Maturity | 3 scene 1 | AI SDK is an ecosystem standard, but RSC generative UI is experimental and discouraged for production. |
| Openness | 4 scene 4 |
Vercel AI SDK generative UI (RSC streamUI, AI SDK UI, @ai-sdk/tui) scene now scene planned
how agents use it
Wiring it in
Low for Next.js/React apps.
Seeing the result
AI state/messages only; the model never sees the rendered UI.
Small edits
createStreamableUI .update/.append/.done on one streamed node; coarse.
History
Message history only.
In short
Agent picks components via tools; no live control of the UI.
architecture
AI SDK RSC (@ai-sdk/rsc): streamUI calls a model and lets tools return React Server Components that stream to the client; createStreamableUI/createStreamableValue push UI or values; useUIState/useAIState separate visual state from model context. The docs say RSC is experimental and recommend AI SDK UI (useChat with typed tool parts rendered by client components) for production. @ai-sdk/tui's runAgentTUI runs a ToolLoopAgent (or a remote ChatTransport) in a ready-made terminal chat UI with collapsible tool/reasoning sections and token statistics.
- Generative UI = tool call -> app-defined component; UI is React code, not data
- RSC path tied to frameworks supporting React Server Components (Next.js)
- runAgentTUI is a fixed chat UI, not a generative or agent-editable surface
- No readback, addressing or history of rendered UI
performance
No published generative-UI performance numbers.
adoption and upkeep
Adoption
The AI SDK is the TypeScript standard for LLM apps; its RSC generative-UI API is a small, experimental corner.
Used by: json-render (uses AI SDK streaming), Vercel templates
Maintainability
Very well maintained SDK; RSC module kept but officially experimental.
Releases: Multiple releases per day across packages · Contributors: ~730 (GitHub contributors API incl. anonymous) · Recent: 1,477 commits on main since 2026-07-11 · Bus factor: high
weaknesses
- RSC generative UI is labelled experimental; docs recommend migrating to AI SDK UI
- UI is code: no serializable UI, no readback, no patching by id
- runAgentTUI is a canned chat interface with display options only
- Tied to React (and RSC-capable frameworks for streamUI)
and thc-scene
Overlap
Agents surfacing UI to users; a terminal front end for agents (runAgentTUI).
What scene would be reinventing
Nothing substantive.
The gap it leaves
Everything scene is about: UI as data, live agent control, readback, patches, data streams, pixels, time travel.
What to borrow
- runAgentTUI's display modes (full/collapsed/auto-collapsed/hidden) for tool and reasoning sections
- ChatTransport abstraction: run local agent or attach to remote one
sources
- AI SDK RSC overview ai-sdk.dev
- runAgentTUI reference ai-sdk.dev
- vercel/ai LICENSE github.com
- npm downloads API api.npmjs.org