Vapi Voice Agent

The Vapi page at /admin/vapi provides programmable AI voice calls with pluggable LLM models and voice providers. Where Bland AI uses its own internal models, Vapi lets you choose the LLM (OpenAI, Anthropic, Groq, Google) and the voice provider (ElevenLabs, PlayHT, Deepgram) for each call.

Architecture

Admin UI → /api/admin/vapi-manage → Vapi REST API
                                     ├── POST /call (create call with transient assistant)
                                     ├── GET /call (list calls)
                                     ├── GET /call/:id (call detail + transcript)
                                     └── GET /phone-number (list Vapi phone numbers)

Calls use a “transient assistant” — the system prompt, model, and voice are defined inline per call rather than referencing a saved assistant. This keeps configuration self-contained in the admin UI.

Page Layout

Stats Row

Four stat cards:

API Status — connection health (OK/Error)
Total Calls — number of calls made
Active — calls currently in progress
Phone Numbers — Vapi phone numbers available for caller ID

Make Call Tab

Form card with:

Customer Phone Number — E.164 format
System Prompt — instructions for the AI assistant (what to say, how to behave)
Model Provider — dropdown: OpenAI, Anthropic, Groq, Google
Model — dynamically filtered by provider selection
Voice Provider — optional: ElevenLabs, PlayHT, Deepgram (or Default)
Voice ID — provider-specific voice identifier
Phone Number ID — optional Vapi phone number for outbound caller ID
Call button with spinner feedback

The model dropdown updates automatically when the provider changes.

Call History Tab

Table of recent calls showing Customer number, Status (badge), Duration, Created date, Cost, and View button.

Call Detail Tab

Full call information:

Customer number, status, duration, timestamps, end reason
Transcript timeline with role-colored message bubbles (amber for assistant, blue for user)
Cost breakdown

API Route

Route: /api/admin/vapi-manage

Auth: Admin-only (CF Access JWT validation)

GET

Health check — verifies Vapi API connectivity.

POST Actions

Action	Parameters	Returns
`get-status`	none	`{ status }`
`make-call`	`customerNumber`, `systemPrompt`, `modelProvider`, `model`, `voiceProvider`, `voiceId`, `phoneNumberId`	`{ success, callId, status }`
`list-calls`	`limit` (max 100, default 50)	`{ calls[], total }`
`get-call`	`callId`	`{ call }` (includes transcript, cost, messages)
`list-phone-numbers`	none	`{ phoneNumbers[] }`

Environment Variables

Variable	Required	Description
`VAPI_API_KEY`	Yes	Vapi API key (from dashboard.vapi.ai)

Model Provider Reference

Provider	Models	Notes
OpenAI	`gpt-4o`, `gpt-4o-mini`	Default provider
Anthropic	`claude-sonnet-4-5-20250929`	Claude Sonnet 4.5
Groq	`llama-3.1-70b-versatile`	Fast inference
Google	`gemini-2.0-flash`	Gemini Flash

Voice Provider Reference

Provider	Voice ID Format	Notes
ElevenLabs	`pNInz6obpgDQGcFmaJgB` (voice hash)	Same voices as ElevenLabs admin page
PlayHT	Provider-specific ID	Alternative TTS
Deepgram	Provider-specific ID	Low-latency TTS

If no voice provider is specified, Vapi uses its default voice.

Vapi Phone Numbers

Outbound calls require a phoneNumberId — a phone number purchased through Vapi’s dashboard. Without one, calls will fail. Check the Phone Numbers stat card or visit dashboard.vapi.ai to purchase a number.

Setup

Create an account at dashboard.vapi.ai
Navigate to Organization → API Keys and copy your key
Set the environment variable:
- Local: add VAPI_API_KEY=your-key to .env
- Production: npx wrangler pages secret put VAPI_API_KEY
Purchase a phone number in Phone Numbers tab ($2/month — required for outbound calls)
Restart the dev server or redeploy

Vapi offers $10 trial credit. Outbound calls cost varies by model and voice provider.

Troubleshooting

API Status shows Error

Symptom: Stats row shows “Error”, all calls fail with 401.

Fix: The VAPI_API_KEY is missing or invalid. Check .env (local) or CF Pages secrets (production). Get your key from dashboard.vapi.ai → Organization → API Keys.

Call fails with “phone number required”

Symptom: Making a call returns an error or never connects.

Fix: Vapi outbound calls require a phoneNumberId — a number purchased through Vapi’s dashboard. Go to dashboard.vapi.ai → Phone Numbers, purchase one ($2/mo), and paste the ID into the “Phone Number ID” field.

AI uses wrong voice or model

Symptom: Call connects but uses unexpected voice/model, or responses seem generic.

Fix: Vapi uses transient (inline) assistant config per call. If Voice Provider is set but Voice ID is blank, Vapi uses its default voice. Get ElevenLabs voice IDs from /admin/elevenlabs or elevenlabs.Izar-Host. Ensure Model Provider matches the Model selected (e.g., OpenAI → gpt-4o).

Cost seems high

Symptom: Individual calls cost more than expected.

Fix: Cost depends on the model and voice provider selected. gpt-4o-mini is cheaper than gpt-4o. ElevenLabs voices add TTS cost on top of the base call cost. Check Call Detail for the cost breakdown. Use the Vapi dashboard for detailed billing.

Twilio Dashboard (/admin/twilio) — manual SMS and voice calls with pre-written text
Bland AI Phone Agent (/admin/bland) — AI calls with Bland’s internal models and voices
ElevenLabs (/admin/elevenlabs) — voice library (ElevenLabs voices can be used as Vapi voice provider)