Poke Voice MCP Server

A stateless TypeScript Model Context Protocol server that lets Poke trigger outbound AI phone calls through Vapi native telephony.

This version is designed for a server-managed Vapi setup: the deployed backend owns the Vapi credentials through environment variables, and Poke clients only provide the call target and assistant instructions.

Highlights

• Streamable HTTP MCP endpoint for web-hosted agents.

• One tool: trigger_outbound_call.

• Vapi-native outbound calls through POST https://api.vapi.ai/call.

• Server-side Vapi credentials only; no API keys in the MCP tool schema.

• Low-cost default model: openai/gpt-4.1-nano.

• Realistic default voice: Vapi Clara, version 2.

• Sanitized structured logs with tracking IDs, user IDs, status, and destination last four digits only.

• Vercel-ready serverless entrypoint.

Related MCP server: Vapi MCP Server

Architecture

Poke
  -> Streamable HTTP MCP request
  -> /api/mcp on Vercel
  -> MCP tool handler
  -> Vapi /call using server env vars
  -> outbound phone call

The backend is stateless per request, but Vapi credentials are deployment-level configuration.

See docs/architecture.md for the security model and request flow.

Tool Contract

trigger_outbound_call

{
  "phoneNumber": "+15551234567",
  "systemPrompt": "You are calling to confirm an appointment. Be concise and polite.",
  "initialMessage": "Hi, this is the appointment assistant calling to confirm your visit."
}

Required fields:

• phoneNumber: destination number in E.164 format.

• systemPrompt: assistant instructions for the call.

Optional field:

• initialMessage: first spoken message from the assistant.

Successful calls return structured JSON:

{
  "ok": true,
  "status": "created",
  "trackingId": "call-id-from-vapi",
  "provider": "vapi",
  "message": "Outbound call initiated."
}

Errors return ok: false with a stable error code, including:

• missing_configuration

• invalid_phone_number

• authentication_failed

• invalid_call_request

• vapi_request_failed

Required Environment Variables

VAPI_API_KEY=your_private_server_side_vapi_key
VAPI_PHONE_NUMBER_ID=your_vapi_phone_number_id
POKE_USER_ID=local-dev-user

POKE_USER_ID is a fallback for local development. In production, Poke can provide x-poke-user-id; the server uses it only for sanitized tracking logs.

Local Development

Requirements:

• Node.js 20+

• Private/server-side Vapi API key

• Vapi phone number ID

Install and build:

npm install
npm run build

Run the HTTP server:

npm run dev

Local endpoints:

GET  http://localhost:3000/health
POST http://localhost:3000/api/mcp
POST http://localhost:3000/api/vapi-webhook
POST http://localhost:3000/api/vapi-inbound

For local testing:

cp .env.example .env

Never commit .env or paste Vapi keys into recipe YAML, prompts, screenshots, logs, or chat messages.

Vercel Deployment

This repo includes:

• api/index.ts: Vercel serverless adapter.

• api/vapi-inbound.ts: Vapi inbound assistant-request handler.

• vercel.json: rewrites for /api/mcp, /mcp, /health, the Vapi webhook, and inbound call handler.

Add production environment variables:

npx vercel env add VAPI_API_KEY production
npx vercel env add VAPI_PHONE_NUMBER_ID production
npx vercel env add POKE_USER_ID production
npx vercel env add VAPI_INBOUND_SYSTEM_PROMPT production
npx vercel env add VAPI_INBOUND_FIRST_MESSAGE production

Deploy:

npx vercel --prod --yes

After deployment, your MCP endpoint is:

https://your-vercel-domain.vercel.app/api/mcp

Update poke.recipe.yaml with your deployed endpoint before publishing the Recipe.

Inbound Vapi Calls

Inbound calls are handled by:

https://your-vercel-domain.vercel.app/api/vapi-inbound

For the current production deployment:

https://poke-voice-rho.vercel.app/api/vapi-inbound

Configure this URL as the Vapi phone number Server URL for:

+13267327987
6aa3552e-1802-46f5-ba58-1a8f1d66b62d

When Vapi sends an assistant-request, the route returns a transient assistant using:

• openai/gpt-4.1-nano

• Vapi voice Clara, version 2

• VAPI_INBOUND_SYSTEM_PROMPT if configured

• VAPI_INBOUND_FIRST_MESSAGE if configured

Manual assistant fallback JSON:

{
  "firstMessage": "Hi, this is Poke Voice. How can I help?",
  "model": {
    "provider": "openai",
    "model": "gpt-4.1-nano",
    "messages": [
      {
        "role": "system",
        "content": "You are a concise, friendly inbound voice assistant for Poke Voice. Answer naturally, ask one question at a time, and help the caller with their request. If you do not know something, say so plainly."
      }
    ]
  },
  "voice": {
    "provider": "vapi",
    "voiceId": "Clara",
    "version": 2
  }
}

## Poke Recipe

The included [poke.recipe.yaml](poke.recipe.yaml) is a template for Poke Kitchen. It declares:

- Streamable HTTP transport.
- No credential setup prompts.
- A confirmation prompt before `trigger_outbound_call` runs.
- Billing and compliance warnings.

Because credentials are server-side, calls bill to the Vapi account configured on the deployment.

## Testing

See [docs/testing.md](docs/testing.md) for MCP smoke tests, live `tools/list` checks, and safe failure-mode tests.

At minimum, run:

```bash
npm run build

Then verify tools/list exposes only call inputs:

• phoneNumber

• systemPrompt

• initialMessage

Compliance

This software can initiate real outbound phone calls. You are responsible for consent, TCPA compliance, local calling laws, Vapi billing, and any platform-specific usage rules.

The project intentionally does not bypass confirmation prompts, rate limits, provider compliance controls, or phone-number ownership requirements.

License

MIT