Tried lets you build AI phone agents that answer inbound calls, place outbound calls and transfer callers to people. You build an agent in the dashboard or through the REST API, attach a phone number, and the agent talks to callers in real time. The API follows Retell's shapes and paths, so code written for Retell runs against Tried with a base URL change.
The default stack is OpenAI gpt-5.6-terra for the language model, a Premium voice, Deepgram for speech recognition, LiveKit for media and Telnyx for telephony. Every model is a hosted API. You can swap each part: 14 text models, 2 speech-to-speech models, 7 voice providers and 5 transcription providers.
How a voice agent works#
Every call runs one agent worker. The worker reads a frozen copy of the agent's configuration when the call starts, then runs a loop for each turn of the conversation.
The caller speaks#
Audio reaches a LiveKit room, from the phone network through SIP or from a browser through WebRTC. Noise cancellation runs first, set by Denoising Mode.
Speech recognition (STT)#
Deepgram nova-3 streams a live transcript. Optimize for speed finalises 25 ms after speech stops; Optimize for accuracy waits 100 ms and formats numbers. See Speech recognition.
Turn-taking decides when the caller is done#
A turn detector model and a voice activity detector decide whether the caller has finished or only paused. Responsiveness sets how long the agent waits. Interruption Sensitivity sets how easily the caller can talk over the agent. See Turn-taking.
The language model (LLM) writes the reply#
The model reads the prompt, the conversation, any knowledge base passages and the available functions. It streams text back, or calls a function such as a transfer or an HTTP request.
Text to speech (TTS) speaks it#
The voice provider turns the first clause into audio while the model is still writing the rest, so the caller hears the start of the reply early.
After the call ends, Tried stores the transcript and recording, runs post-call analysis, bills the call per second and sends webhooks.
How Tried keeps latency low#
The builder header shows an estimated latency range for your agent's current model, voice and transcription settings, and each call records the measured end-to-end latency as percentiles. These techniques are on by default:
| Technique | What it does |
|---|---|
| Preemptive generation | The LLM starts on the interim transcript before the turn is confirmed. Audio is only synthesised once the turn is confirmed, so no speech is wasted. |
| Warm connections | The worker opens its STT, LLM and TTS connections while the phone is ringing, and sends a warm-up request that fills the LLM prompt cache. |
| Turn detector | A model judges whether the sentence is finished, so the agent does not wait the full silence window after a complete sentence. |
| Fast transcription | Speed mode finalises the transcript 25 ms after the caller stops. |
| Streaming TTS | Speech starts from the first clause of the reply. |
| Fast tier and Ultrafast | Optional per agent: priority serving from the model provider. See LLM options. |
Read How it works for the full call lifecycle.
Explore the docs#
Start here#
Popular use cases#
The Create agent dialog ships templates for these jobs. Each template sets a prompt, a voice, starter functions and post-call fields that you then edit.
| Template | What the agent does |
|---|---|
| Front-desk Receptionist | Answers inbound calls, answers FAQs and routes callers to the right person. |
| Appointment Booking | Books, reschedules or cancels appointments and confirms the details. |
| Lead Qualification | Qualifies inbound or outbound leads on budget, authority, need and timeline. |
| Outbound Sales | Calls warm leads, pitches the offer, handles objections and books a demo. |
| Customer Service | Resolves account, order and billing questions, and escalates when needed. |
| Medical Clinic | Books patients in, answers clinic questions and routes urgent symptoms. |
| Real Estate Enquiries | Answers listing enquiries, qualifies buyers and books inspections. |
| Home Services Dispatch | Takes job requests for plumbing, electrical or HVAC work and schedules a technician. |
FAQ#
Is Tried compatible with the Retell API?
Yes for the core surface: agents, Retell LLMs, conversation flows, calls, phone numbers, knowledge bases, batch calls, voices, chat and webhooks use Retell's paths and field names. Webhooks carry both x-tried-signature and x-retell-signature. A few endpoints use different paths. See Migrate from Retell.
Which phone numbers can I use?
You can buy numbers through the API from Telnyx or Twilio, or connect a number you already own on Twilio or any SIP trunk. Agents answer, place and transfer calls on Australian numbers. See Telephony.
Can I use my own language model?
Yes. Pick from the built-in OpenAI, Anthropic and Google models, or create a Custom LLM agent that connects to your WebSocket server and streams the replies itself. See Choose an agent type.
Do I need my own model or voice API keys?
Not to build and test. Each provider runs in mock mode until its key is set on the platform, so agents, the playground and the call lifecycle work end to end without any keys. Real calls, voices and transcription need the provider keys configured on the deployment.
How is a call billed?
Per second, from a per-minute rate made of voice infrastructure, the LLM, the voice and telephony, plus add-ons such as knowledge base retrieval or guardrails. The builder header shows the rate for your agent's current settings under Cost.
Can the agent speak languages other than English?
Yes. The Language picker offers 73 locale codes, including multi. Select more than one language under Languages callers may speak and transcription switches to multi-language mode. See Languages.
Can I build a text chat agent too?
Yes. Chat agents use the same prompt or flow engines as voice agents and run over the chat API, the website widget and SMS. See Chat agents.