Skip to main content
vapi.health
All guides
Vapi basics

Vapi Dashboard Guide: Assistants, Logs & Setup

A practical tour of the Vapi dashboard: where each setting lives, which ones actually change call quality, and how to read a call log when an agent misbehaves.

Last updated September 3, 2026

Dashboard layout in one paragraph

The Vapi dashboard is organized around five things: Assistants (the configuration of a voice agent), Phone Numbers (inbound and outbound telephony, including SIP trunks), Call Logs (per-call transcripts, recordings, latency, and tool traces), Tools and Files (function definitions and knowledge sources), and Account settings (API keys, provider keys, org members, usage, and billing). Nearly everything you can click in the dashboard is also an API field, so treat the UI as a fast editor for a JSON assistant object rather than as the source of truth.

Configuring an assistant

  • Model: pick the LLM and paste the system prompt; keep it under roughly 800 tokens for latency.
  • Transcriber: choose the STT provider, language, and endpointing/wait-time settings.
  • Voice: choose the TTS provider and voice ID, and prefer the streaming variant.
  • First message: set the greeting so the caller hears audio immediately.
  • Tools: attach functions with strict JSON schemas and short descriptions.
  • Advanced: silence timeout, max duration, interruption sensitivity, background-sound settings.

Choosing STT, LLM, and TTS in the UI

The three provider dropdowns are the highest-leverage controls in the dashboard. On the transcriber, a streaming provider with interim results plus an endpointing value in the 200–300ms range is the difference between a snappy agent and one that talks over people. On the model, a small fast LLM on the hot path beats a large one for conversational feel; route only complex tool decisions to a bigger model. On voice, streaming TTS with a pre-warmed voice cuts hundreds of milliseconds off time-to-first-audio. Change one dropdown at a time and re-test with the web widget so you can attribute the difference.

Reading call logs and the latency waterfall

Every call in the log expands into a turn-by-turn timeline: user audio, transcript finalization, LLM time-to-first-token, TTS time-to-first-byte, and playback start. This waterfall tells you exactly which stage is slow, which is the only reliable way to debug 'the agent feels sluggish'. Look for three classic shapes: a long gap before transcript finalization (endpointing too conservative), a long LLM segment (prompt too big or model too large), and a long TTS segment (non-streaming voice or a cold voice). The same view shows tool-call payloads and responses, which is where most wrong-answer bugs actually live.

Phone numbers and SIP trunking

You can buy a number inside Vapi, import a Twilio or Vonage number with credentials, or connect your own carrier over SIP for enterprise routing. Each number gets an inbound assistant assignment and can be used as a caller ID for outbound calls. For SIP, you configure the trunk's URI and credentials and, if the carrier requires it, an allowlist of Vapi's signaling IPs. Test inbound and outbound separately — a trunk that accepts inbound calls can still fail outbound on authentication or codec negotiation.

Webhooks and server URLs

The server URL is where Vapi posts events: call started, transcript updates, tool calls, end-of-call reports, and status changes. Set it per assistant or globally, add the shared secret so your endpoint can verify requests, and subscribe only to the events you handle. Two rules keep production stable: respond to tool calls in under 300ms, because the caller is waiting in silence, and return HTTP 200 quickly on informational events, doing heavy work asynchronously.

Usage, billing, and provider keys

The usage view breaks spend into Vapi orchestration minutes and pass-through provider costs, which is the fastest way to spot a runaway configuration such as an expensive TTS voice on high-volume calls. Provider keys live in account settings: adding your own OpenAI, Deepgram, or ElevenLabs keys switches those line items to direct billing at provider rates. Set a spend alert before you launch any outbound campaign — the failure mode of a retry loop is a very large bill rather than an error message.

Troubleshooting the dashboard itself

  • Dashboard loads but calls fail: check provider keys and quota before blaming Vapi.
  • Login or session errors: clear the session, retry SSO, and confirm you are in the right organization.
  • Blank call logs: logs are scoped per org and per API key — verify the key that placed the call.
  • Changes not taking effect: assistants are versioned per call; a call already in progress keeps its old config.
  • Dashboard unreachable: check the platform status page and this site's live reachability checks before rolling back your own deploy.

FAQ

How do I create and test an assistant in the Vapi dashboard?
Navigate to the Assistants tab in the Vapi dashboard, configure your system prompt, select STT, LLM, and TTS providers, and click Test to speak directly with your voice agent in the browser playground.
Where can I find call logs and latency metrics in Vapi?
Go to the Call Logs section of the Vapi dashboard to review audio recordings, end-to-end transcripts, cost breakdowns, and granular STT, LLM, and TTS latency waterfalls for every conversation.
How do I connect custom telephony numbers to Vapi?
In the Phone Numbers tab, import credentials from Twilio, Vonage, or Telnyx, or purchase pre-configured numbers directly within the Vapi dashboard to assign incoming and outgoing call assistants.
Where do I configure a Vapi assistant's model and voice?
Open Assistants in the dashboard, select the assistant, then use the Model, Transcriber, and Voice panels. Model holds the LLM choice and system prompt, Transcriber holds the STT provider and endpointing settings, and Voice holds the TTS provider and voice ID. Every field maps to the assistant object in the API, so you can move the same configuration into code later.
How do I see why a Vapi call was slow?
Open the call in Call Logs and read the per-turn latency waterfall. It separates transcript finalization, LLM time-to-first-token, TTS time-to-first-byte, and playback start, so you can see which stage dominated. Long pre-transcript gaps mean endpointing is too conservative, long LLM segments mean the prompt or model is too big, and long TTS segments usually mean a non-streaming voice.
Can I add my own provider API keys in the dashboard?
Yes. Account settings has a provider keys section where you can add your own OpenAI, Anthropic, Deepgram, ElevenLabs, Cartesia, and telephony credentials. Once added, those services bill you directly at provider rates instead of being marked up through Vapi, while Vapi continues to charge its per-minute orchestration fee for the call itself.
How do I connect a phone number or SIP trunk?
Use the Phone Numbers section to buy a number from Vapi, import an existing Twilio or Vonage number with its credentials, or configure a BYOC SIP trunk with its URI and authentication. Then assign an inbound assistant to the number. Test inbound and outbound paths separately, since authentication and codec problems usually surface on only one direction.
Why are my assistant changes not affecting a live call?
Vapi snapshots the assistant configuration when a call starts, so an in-progress call keeps the settings it began with. Place a new test call after saving. If a new call still uses old behavior, confirm you edited the assistant that the number or API request actually references — duplicated assistants with similar names are the most common cause.
Can I test an agent without a phone number?
Yes. The dashboard includes a browser talk-to-assistant widget that runs the same pipeline over WebRTC, so you can iterate on prompts, voices, and endpointing without buying a number. Browser calls skip telephony codec overhead, so measure final latency on a real phone call before you promise a number to stakeholders.
How do I monitor Vapi spend?
The usage and billing views split cost into Vapi orchestration minutes and pass-through provider charges, filterable by date. Check it after every configuration change that touches the voice or model, set a spend alert before launching outbound campaigns, and watch for retry loops, which show up as a spike in very short calls rather than as an error.
What should I check first when the Vapi dashboard seems down?
Confirm whether the problem is the dashboard, the API, or your own provider keys. Check Vapi's official status page and the live reachability checks on this site, then try an API call directly with curl. If the API answers while the dashboard does not, your calls are probably still running and the issue is limited to the console.

Keep reading

Vapi basics
How Vapi Voice Agents Work: Architecture & Pipeline
Vapi basics
Vapi Pricing Per Minute: 2026 Cost Breakdown
Compare platforms
Vapi vs Retell vs Bland: Voice AI Comparison (2026)

Related on this site