Agent
Define model and tool-loop agents for Vite and Nitro apps.
@vitehub/agent defines server-side agents. An agent owns model instructions, model-facing capabilities, and the run or stream call that produces a response. It consumes @vitehub/runtime capabilities when it needs host resources.
Use Agent when a server feature needs a model loop with ViteHub message input.
server/agents/triager.ts
import { defineAgent, type AgentToolDefinition } from '@vitehub/agent'
import { getMessageText } from '@vitehub/agent'
const classifyTicket: AgentToolDefinition<{ message: string }, { queue: string; priority: string }> = {
name: 'classifyTicket',
description: 'Classify a support request before queue handoff.',
policy: ({ input }) => {
const message = typeof input === 'object' && input && 'message' in input
? String(input.message)
: ''
return /refund|invoice|payment/i.test(message) ? 'require-approval' : 'allow'
},
}
export default defineAgent({
description: 'Triage support requests and prepare a queue handoff.',
async run({ input }) {
const latest = input.messages?.at(-1)
const message = latest ? getMessageText(latest) : ''
return {
raw: { tool: classifyTicket.name },
text: `Classify and route: ${message}`,
}
},
})
What Agent owns
Agent definitions
Keep model instructions, capabilities, and custom run behavior in
defineAgent().Message input
Accept
@vitehub/agent input and convert it to model calls inside Agent.Optional routes
Expose discovered agents over HTTP only when you enable generated routes.
What Agent does not own
Agent does not own chat webhooks, Chat SDK adapters, workflow runs, runtime capability registration, or sandbox lifecycle. Use the package that owns each boundary:
| Need | Use |
|---|---|
| Receive Slack, Discord, Telegram, or Teams events | @vitehub/agent/chat |
| Store or replay conversation state | @vitehub/agent |
| Resolve shared capabilities, approvals, and trace context | @vitehub/runtime |
| Coordinate durable work | @vitehub/workflow |
| Execute isolated code | @vitehub/sandbox |