AI agents · Foundations live, Agents API coming

Agents that actually
do the work.

Concierges that answer guest questions at 3am. Ops bots that triage messages and dispatch tasks. Voice agents that confirm a check-in. Build them on the Repull data plane today, with the MCP server, the AI SDK provider, and — soon — the Repull AI Agents API.

What you build

Agents that read, write, and decide.

The foundations are in place today. Eighteen MCP tools and a Vercel AI SDK provider mean a model can hit Repull's data plane natively, without you wiring up a custom function-calling layer.

Guest concierge

Wi-Fi password, parking instructions, restaurant suggestions, late check-out request handling. Pulls from your KB, falls back to the host, escalates the right things.

Ops bot

Triages inbound messages, classifies urgency, dispatches to the right team, opens tasks, follows up. Reads reservations, listings, and guest history natively.

Voice agent

Inbound calls — confirm a check-in, give a code, escalate. The model has the same tools the chat concierge does. Carrier integration handled by your provider of choice.

Available today

Plug Repull into any AI SDK chat.

@repull/ai-sdk is a Vercel AI SDK provider. Pass it to streamText and the model gets a tool surface for reservations, listings, messaging, and pricing — typed, scoped to your workspace, ready to call.

  1. 1

    Install the provider

    It's a thin wrapper around the SDK that exposes Repull operations as tools. Works with any AI SDK model — OpenAI, Anthropic, Google, the rest.

  2. 2

    Drop it into your chat

    Pass repullTools to streamText and the model can call them. No custom function-calling layer.

  3. 3

    Scope keys per agent

    Each agent gets its own scoped API key. Read-only for a concierge, read+write for an ops bot. The agent never sees more than it needs to.

agent.ts · ts
import { streamText } from 'ai'
import { openai } from '@ai-sdk/openai'
import { repullTools } from '@repull/ai-sdk'

const tools = repullTools({
  apiKey: process.env.REPULL_API_KEY!,
  workspaceId: process.env.REPULL_WORKSPACE_ID!,
  scope: 'concierge', // 'concierge' | 'ops' | 'admin'
})

const result = streamText({
  model: openai('gpt-4o'),
  system: "You are a vacation rental guest concierge. Answer in the guest's language. Escalate refunds, medical, and legal matters to a human.",
  messages,
  tools,
})
claude_desktop_config.json · json
{
  "mcpServers": {
    "repull": {
      "command": "npx",
      "args": ["@repull/mcp"],
      "env": {
        "REPULL_API_KEY": "sk_test_...",
        "REPULL_WORKSPACE_ID": "ws_..."
      }
    }
  }
}

Same shape works in Cursor, Windsurf, or any MCP-aware client. The MCP server ships with 18 tools today.

Coming

The Repull AI Agents API.

Above the SDK provider sits the Agents API — pre-built guest-experience agents with the prompt engineering, evals, and audit trail already done. Drop in your KB, point a phone number or chat channel at it, and it's live. The same agents Vanio AI runs in production.

  • Guest concierge agent — chat + voice
  • Ops triage agent — inbound message routing
  • Pricing copilot — explains the recommendation
  • Review responder — drafts on every new review
  • Custom agents — your prompt, our tools and evals

Behind the scenes: the same agent runtime that powers Vanio AI's host-facing concierge in production. Tested against thousands of guest conversations a week.

Get on the Agents API design list

What kind of agent are you trying to ship? The earliest design partners get the API surface they ask for.

Goes straight to the team. No list, no drip.

Send

Build the agent today.

The data plane is already live. The MCP server and AI SDK provider are already shipping.