The shortest path to a working voice agent, entirely in the dashboard: add the provider keys it runs on, write a prompt, talk to it in your browser, and publish it. No phone number and no code. For a real phone line afterwards, see the phone number quickstart.
1

Sign in

Go to https://app.talqing.com and sign in with Google. Your first sign-in creates a workspace with $10 of credit for the platform fee.
2

Add provider keys

Talqing runs every model on your provider accounts, so nothing works until you add keys. A voice agent needs a language model, speech-to-text and text-to-speech. Two keys cover all three:Go to Build → Provider Keys and add both. Each key is checked with the provider before it is saved. Only an ADMIN can add keys; if you created the workspace, that is you. See provider keys for every provider.
3

Create the agent

Build → Agents → New agent, give it a name unique in the workspace, and leave Channel on voice. A new agent is a draft: it reaches no caller until you publish.
4

Write the prompt and greeting

In Prompt & voice, set the system prompt:
And the greeting, spoken the moment the call connects:
A length limit, speech-friendly formatting and an explicit “never guess” are what make a voice prompt work. See prompting.
5

Pick the models

In Models, keep the pipeline on cascade (speech-to-text → LLM → text-to-speech) and choose models your keys cover:The right rail shows a live per-minute cost estimate, split into provider spend and the platform fee. See models for the full catalog and realtime for speech-to-speech models. Save draft.
6

Talk to it

Scroll to Test call and press Start call, then allow the microphone. The test panel runs the draft, saving your edits first, and validates it the way publishing does, so a missing provider key is refused here with a message such as no Deepgram API key for this workspace - add one under Provider keys.Try interrupting the greeting, asking for the hours, and asking something it cannot know (“how much is a root canal?”) to check it offers a message instead of inventing an answer.
A test call is a real call: it spends your provider keys, is charged the platform fee, and is recorded by default.
7

Publish

Press Publish. This validates the draft and freezes it as v1, the version that phone numbers, batches and API calls run. Errors block a publish; warnings are shown and do not. Later edits change only the draft until you publish again. See versions and validation.
8

Read the call back

Open Monitor → Calls and select your test call. It shows the transcript with every tool call, latency per stage (speech-to-text, LLM, text-to-speech), cost split into provider spend and platform fee, the stereo recording, and the exact agent configuration the call ran. See calls.

​
Next

Take a real phone call

Import a number from your carrier and point this agent at it.

Give it a tool

Look something up, book a slot, transfer to a human.

Write a better prompt

Voice-specific prompt craft.