What happens if a prompt change makes things worse?

Every save creates a new revision, and you can restore any earlier one with one click. If you want to see exactly what changed, compare shows a field-level diff between two agents, so you are not guessing which line broke the call.

Can a call move between more than one configured agent?

Yes. A squad hands a caller from one agent to another inside the same room, for example from a general line to a billing specialist, and carries the transcript with it. The caller does not start over with a stranger.

Which models and voices can I use?

Speech-to-text runs on Deepgram by default, with OpenAI Whisper and Cartesia also supported. The language model is OpenAI by default, on a model-agnostic architecture that also supports Anthropic, Google, and OpenRouter. Text-to-speech runs on ElevenLabs, Cartesia, or FishAudio. Provider keys are your own, encrypted per organization and validated before they go live.

How do I test a change before a caller hears it?

An AI prompt builder can audit a prompt and suggest edits before you save it. Text-mode testing runs the same agent as a chat conversation, with every tool call it makes visible in the transcript, so you can see what it would have done on a real line.

Build → Agents

AI voice agents that answer like they work for you.

Configure the prompt, voice, model, tools, and knowledge from a dashboard, not a deploy. The database is the single source of truth, fetched fresh on every call, and every edit leaves a revision you can restore.

Configured in a dashboard, not a deploy
One-click restore to any revision
Squads hand off without losing the thread
Tested in text before it takes a call

Book a demo See a call move through it

Greeting starts, cached: ~30ms
First real reply, after the warm-cache ping: 4.5s → 1.5–2.5s
Prompt-cache hit ratio, steady state: 0.98–0.99

Measured in production for a single flagship operator. Your first-call time depends on prompt length and provider.

Agent configuration → live call traceNo redeploy

Agent configuration

Core
Identity, prompt, LLM
Voice
Voice, language, STT
Capabilities
Tools, MCP, transfer
Tuning
Timing, audio
Channel
Email, chat config

Live call trace

STT chunk arrivesstreaming
Deepgram, nova-3:multi
LLM calls a toolcache warm
check_availability()
TTS starts speakingchunk queued
ElevenLabs, eleven_turbo_v2_5

Fetched fresh from the database on every call. Edit a group in the dashboard and the next ring uses it. No redeploy.

Config to call

Save it in the dashboard. The next caller hears it, live.

Every agent pulls its own configuration from the database at the start of the call. Nothing is baked into a deployed build, so nothing about your agent is stuck waiting on an engineering sprint.

A business owner on the phone at his counter while a second call comes in on another line

He is already on one call. The next one goes to voicemail.

On the call

Talk to it like a person. It keeps up.

A voice that sounds good is the easy part. These are the mechanisms working under every call, not settings you have to go find.

It knows the difference between you and the TV.
A real interruption has to clear a few checks first. Is this a voice. Is it long enough. Is it not background chatter. Only then does the agent stop talking and listen, so a television or a radio in the room will not cut it off mid-sentence.
It waits while you read out the numbers.
A purpose-built model tracks whether you are actually still speaking, not just whether you paused. It waits through a full account number, phone number, or card number, digit by digit, then replies the instant you are done. No mid-number cut-ins, no rushed guesses.
It talks while it works, in your language.
Before it runs a lookup, the agent says something natural first, in the caller's own language. A lookup never sounds like the line went quiet. No dead air while it checks an order or an account.
It briefs the human before it hands off.
On a warm transfer, the caller goes on hold while the agent briefs the person picking up. Then it either connects the call or comes back to the caller with an answer. Never a blind handoff.
See how transfers work
It cleans up the line before it starts listening.
Real-time speech enhancement and speaker isolation strip background noise and crosstalk out of the audio first. The agent hears the caller, not the room they are calling from.
Shipped
It finishes the sentence before it acts.
A transfer or a hangup waits behind the agent's own last sentence. It says goodbye, or finishes briefing the human, before the action fires. No caller is cut off mid-sentence.

The agent isn't just a talker. It's an operator on the line.

end_call
warm_transfer
cold_transfer
hold / resume
transfer_to_agent
book_meeting
crm.write
+ your tools

Every one of these is a tool call the agent can decide to make mid-call, logged like anything else it does.

Build

Four ways the agent stays yours after launch

Most voice bots ship once and calcify. Here, you ship an agent the way you ship code: branch a draft, diff it against production, canary it on a slice of calls, then merge it in.

The dashboard is the source of truth

Prompt, voice, model, tools, and knowledge live in the database, not in a deployed build. An agent fetches its own configuration at the start of every call, so a change you make now is live on the next ring, not the next release.

No redeploy for a prompt, voice, or model change
The same configuration path for voice, chat, and email
Five section groups: Core, Voice, Capabilities, Tuning, Channel

Agent editor

CoreIdentity, prompt, LLM
VoiceVoice, language, STT
CapabilitiesTools, MCP, transfer
TuningTiming, audio
ChannelEmail, chat config

Fetched by the agent process at the start of every call. Nothing here waits on a build.

Branch a draft, diff it, then merge it in

Every edit leaves a revision you can restore in one click, and compare shows a field-level diff between any two agents. Roll a new version out to a slice of live calls first, a canary, before it takes every call. If it is not right, roll back in one click.

Full agent revision history, one-click restore
Prompt version history: manual, AI-suggested, or reverted
A field-level compare (diff) between any two agents
Canary a new version on a slice of live traffic before it takes every call

Revision history

Rev 12

Rev 13

Rev 14

Rev 14 · canary · 12% of calls

Restored to Rev 12 in one click. A canary rolls back the same way.

Product interface · sample data

One caller, more than one specialist

A squad hands a caller from one configured agent to another inside the same room, general line to billing, sales to support, and carries the transcript along. The caller never repeats the story to a second bot.

Handoff can be silent or announced to the caller
The full transcript is preserved across the handoff
Each specialist keeps its own prompt, voice, and tools

Squad handoff

Sales agentBilling agent

Transcript carried over. The billing agent opens already knowing what the caller asked for.

Test it in a live playground. Then listen in for real.

An AI prompt builder can audit a prompt and suggest edits before you save. A live text playground runs the same agent as a real conversation, with every tool call it makes visible, so you see what it would have done before a caller does. Once it is live, a supervisor can listen in on a real call without interrupting it.

AI-assisted prompt audit and suggestions
A live text playground with full tool-call visibility, before a caller hears it
Export a test run to JSON or CSV
Supervisor listen-in on any live call

Live text playground

Tester: Can I move Thursday's pickup?

Agent: Let me check that slot for you.

tool_call: check_availability → ok

Supervisor listening, silently

Every tool call is visible before a caller ever hears the change. Once it is live, a supervisor can drop in on any call the same way.

Providers

Choose the models and voices, not the other way around

Speech-to-text, the language model, and text-to-speech are each a setting, not a lock-in. Swap a provider without migrating the agent.

Speech to text

Deepgram (default, nova-3)
OpenAI Whisper
Cartesia

Language model

OpenAI (default)
Anthropic
Google
OpenRouter

A model-agnostic architecture. OpenAI carries the full production feature set today; Anthropic, Google, and OpenRouter are supported alongside it.

Text to speech

ElevenLabs
Cartesia
FishAudio

Bring your own provider keys. Each one is encrypted per organization and validated with a live ping before it goes into service.

Measured, not modeled

86%

Interruption-detection precision, measured against real call audio, not a benchmark set.

11: Named latency optimizations between a caller's last word and the reply
2: Languages live today, English and Spanish, on a stack built for more

Every number here is a query we can rerun, not an estimate we are hoping holds up.

Questions

What teams ask before they change an agent

Still have questions?

Bring a prompt you are not happy with. We will show you the revision it leaves.

Book a demo Talk to sales

Everything lives in the dashboard: identity and prompt under Core, voice and speech-to-text under Voice, tools and transfer rules under Capabilities, timing and audio under Tuning, and channel settings under Channel. Save a change and the next call reads it straight from the database. There is no build to redeploy.

Book a demo

Configure it once. Watch it answer.

Bring your own prompt or start from a template. We will set up an agent with your voice and your knowledge, live, and show you the revision it leaves behind.