@uservane/vercel-ai

Agent-native capture for chat UIs built on the Vercel AI SDK. Headless AgentController plus React <UserVaneProvider> / <InlineFeedback />. Reuses @uservane/browser for bootstrap, show-decision, and submit.

Install

npm install @uservane/vercel-ai
# peers: react >= 18; ai >= 3 (optional peer)

Client exports

Export	Role
UserVaneProvider	Init controller, optional identity, default InlineFeedback
useUserVane()	resolveTask, bind, isTaskOpen, identify, survey, dismiss, complete, noteUserTurn, active, controller
InlineFeedback	Inline ask UI (0 CLS, dismiss-on-next-action)
AgentController	Headless controller class
UserVaneErrorBoundary	Error boundary
Correlation helpers	explicitCorrelationSource, emptyCorrelationSource, …

Provider props (selected)

• apiKey (publishable)
• apiBase, debug
• nonBlocking (default false: blocking-capable, task open until resolveTask)
• surveySlug default for resolve / mechanical triggers
• userId, traits
• correlationSource (prefer bind for production linked path)
• renderInline (default true)

Controller methods (selected)

Method	Role
init / identify	Publishable-key setup and respondent identity
resolveTask(opts?)	Mark task resolved; run show-decision
isTaskOpen()	Open-task bar query
bind(binding)	Install server-minted { surveyId, showToken, sessionId }
unbind(surveyId?)	Clear binding(s)
complete / dismiss	Submit or dismiss active ask
survey(slug)	Imperative trigger
noteUserTurn	Re-open task after resolve in blocking mode; advances afterTurn when nonBlocking
afterTurn / onSessionEnd	Mechanical triggers (nonBlocking only)
setHeadlessContract	Required pair isTaskOpen + onDismiss when headlessMode

Server: mintFeedbackToken

import { mintFeedbackToken } from "@uservane/vercel-ai/server";

const { tokens, suppression, issuedAt } = await mintFeedbackToken({
  secretKey: process.env.USERVANE_SECRET_KEY!,
  respondentId: "u_123",
  sessionId: "sess_abc",
  surveyId: "surv_post_task", // optional
  taskType: "refund", // optional, untrusted display metadata
});

• Calls POST /v1/sdk/tokens with the secret key.
• Never import @uservane/vercel-ai/server from client code.

Linked path (summary)

1. Server: same sessionId as Langfuse attributes → mintFeedbackToken
2. Client: bind({ surveyId, showToken, sessionId })
3. Client: resolveTask() at true outcome → InlineFeedback / headless complete
4. Optional server: @uservane/langfuse-push

Agent quickstart · TypeDoc client · TypeDoc server