Voice providers

Credentials and supported features for the voice providers Simulation connects to

Simulation places voice calls through a connected provider. Vapi, Retell, Bland.ai, and LiveKit are the native providers, chosen when you set up a voice agent definition. Others covers any agent you can reach by phone.

Vapi

Connecting a voice agent to Vapi needs the API key and the assistant ID from your Vapi account, both entered when you connect your agent. If the agent takes inbound calls, it also needs a contact number of 10 to 12 digits.

Retell

Connecting a voice agent to Retell needs the API key and the assistant ID from your Retell account, entered the same way as Vapi. Inbound calling needs the same 10 to 12 digit contact number.

Note

Retell agents don’t support tool call evaluation. The “Enable tool call evaluation” toggle on a run has no effect on Retell-backed calls; evaluation covers the conversation only.

Bland.ai

Bland.ai appears as Bland.ai in the provider dropdown and authenticates with an API key, the same as Vapi and Retell. Two things work differently:

  • The Assistant ID field takes a Conversational Pathway ID. Bland has no separate assistant object, so open the pathway your agent runs and copy its ID into Assistant ID
  • A contact number is always required, not just for inbound. Bland has no web connector, so every simulated call is placed over the phone. For inbound tests, use a Bland number attached to that pathway

Note

Bland records a single combined audio track rather than separate caller and agent channels. An eval whose input is mapped to the stereo, assistant, or customer recording resolves empty on a Bland call, so map whole-conversation evals to call.voice_recording instead. See Edit a run’s evals for where that mapping lives.

Because Bland’s API key is sent in a different header format from the other providers, switching an existing agent definition away from Bland clears the stored key. Re-enter it for the new provider.

LiveKit

Choose LiveKit for an agent that runs as a LiveKit agent worker. It connects with your LiveKit server details instead of an API key and an Assistant ID:

  • LiveKit Server URL: your LiveKit server, such as https://your-project.livekit.cloud. A wss:// URL works too
  • LiveKit API Key and LiveKit API Secret
  • Agent Name: the name your worker registers with WorkerOptions(agent_name=...). It must match exactly; a display name won’t work

Click Test Connection. You can’t move to the next step until the connection succeeds. Room Config JSON is optional.

Under Contact Information, choose how test calls reach the agent:

  • Web: for each call, Simulation creates a room on your LiveKit server and sends your agent into it by its Agent Name. No phone call is placed and no number is needed
  • Phone: Simulation places a real phone call to the number you enter. Use a number that reaches your agent, and leave Inbound Calls on: with it off, Simulation doesn’t dial the number and waits for your agent to place the call instead. A phone call doesn’t go through your LiveKit server, and phone tests run one call at a time

Agent speaks first decides who opens the call. Turn it on if your agent greets first, and the simulated caller waits for the greeting. With it off, the simulated caller opens.

Max Concurrent Sessions caps how many web test calls run against your agent at once. It starts at 5.

Note

On a self-hosted install, Phone and Agent speaks first need the hosted simulation runner with voice turned on (HOSTED_RUNNER_ENABLED and HOSTED_RUNNER_VOICE_ENABLED). Phone also needs the install’s own LiveKit server and an outbound SIP trunk (LIVEKIT_OUTBOUND_TRUNK_ID and PSTN_CALLER_NUMBER). Without the runner, LiveKit calls always go over the web, even with a number saved, and Inbound Calls decides who speaks first: on, the simulated caller opens; off, your agent does.

LiveKit doesn’t share your agent’s prompt, so nothing syncs from it. Paste the system prompt into Prompt / Chains on the Behaviour step.

What each provider supports

CapabilityVapiRetellBland.aiLiveKit
Voice callsYesYesYesYes, over the web or by phone
Call recordingsYesYesYes, one combined trackYes
Call transcriptsYesYesYesYes
Compare with baseline (replay)Yes, the original call’s configuration comes acrossTranscript onlyTranscript onlyNo, Observe has no LiveKit call history to replay
Tool call evaluationYesNoNoNo
Contact numberInbound onlyInbound onlyAlways requiredOnly for Phone

Recordings and transcripts show up in the call detail view the same way regardless of provider. Vapi, Retell, and Bland.ai support “Compare with baseline” on a call, which lines its transcript up against the reference call. What differs is how much of the original comes across, and Replay is where that difference is explained.

Limits

Voice calls are capped at 30 minutes on any provider; a call in progress is ended automatically once it hits that mark. See Call metrics for how duration, cost and talk time are reported once a call completes.

Was this page helpful?

Questions & Discussion