This is the shortest path to a working voice agent: sign in, add the provider keys it will run on, write a prompt, publish, and talk to it in your browser. Fifteen minutes, no phone number and no code. At the end you will read the call back — the transcript, where the time went, and what it cost.
1

Sign in

Go to https://app.talqing.com and sign in with Google. That is the only way in — there is no password to set.Your first sign-in creates a workspace for you and grants it $10 of credit in every region. That is the platform fee for roughly 2,800 minutes of voice, so the Talqing side of everything below costs you nothing today.The dashboard opens on one region. It does not matter which for a browser test; if you are in India, in is closer and the agent will feel faster. See regions.
2

Add your provider keys

Nothing runs until you do this. Talqing holds no provider credentials of its own — every model your agent uses is called with your key, on your account, and those providers bill you directly. An agent whose provider has no key is refused at publish, naming the provider.You need a key for three things: a language model, speech-to-text and text-to-speech. Go to Build → BYOK in the sidebar and add:Two keys cover all three slots, which is why this pairing is a good place to start. Any combination from the ten supported providers works; you can change your mind later.Each key is proved against the provider before it is stored. A key the provider rejects is a 400 and nothing is written; a provider that does not answer is a 502 and nothing is written either. So a bad paste cannot break a workspace that was working.Only an ADMIN can add keys. If you created this workspace, that is you.
3

Create the agent

Build → Agents → New agent. Give it a name — it has to be unique in the workspace — and it opens in the editor.A new agent is a draft. Drafts run nothing and reach nobody. The editor is one long scroll with a sticky section nav down the side, and a toolbar at the top showing draft, unsaved when the editor is ahead of the server, and later v1 live.
4

Choose the voice channel

The Channel section at the top offers voice, video and text. Leave it on voice.The channel decides which model slots exist. Voice gets speech-to-text, a language model and text-to-speech; text gets only a language model; video adds an avatar. See channels.
5

Write the prompt and the greeting

In Prompt & voice, paste this into the system prompt. It is a real starting point, not a placeholder — change the name and the facts and it will run.
Then the greeting — the first line spoken the moment the call connects, before the model has done anything:
Three things about that prompt are doing real work, and they are the difference between an agent that sounds like a phone call and one that sounds like a chatbot reading aloud:
  • A length instruction. Without one, a language model writes paragraphs, and a paragraph read at speaking speed is fifteen seconds the caller cannot interrupt gracefully.
  • A speech instruction. 09:00-19:00 is a string a text-to-speech engine will read as digits and punctuation.
  • A refusal instruction. “Never guess” is the single highest-value sentence in a voice prompt. A model that invents a price on a recorded line is a problem you find out about later.
Voice-specific prompt craft has a page of its own: prompting.Leave the rest of this section alone for now. Save draft.
6

Pick the models

Scroll to Models. Leave the pipeline on cascade — speech-to-text, then a language model, then text-to-speech, three swappable parts. (Realtime is one speech-to-speech model instead, and gives up some control in exchange for latency.)Set the three stages to models your keys cover:Every model in the picker comes from the catalog, and each entry compiles to that provider’s own API — nothing is routed through a broker.The right-hand rail shows a per-minute cost estimate for the draft as you change it, split into speech-to-text, language model, text-to-speech and the platform fee. For this stack it lands near four cents a minute, of which $0.0035 is Talqing’s. Use it to compare two stacks before you commit; see pricing.Save draft.
7

Publish

Press Publish. This freezes the draft as v1 and makes it the version that runs.Publishing is also the gate that checks everything: a model that does not exist, a template that reads a variable nothing supplies, and — the one that catches most people — a provider with no key:
That check runs at publish and never on a draft save, so a draft that saves cleanly is not proof it can be published. Errors block a publish; warnings do not and come back on the response. See validation.From here on, editing the agent changes the draft. The toolbar tells you whether the draft differs from v1, and nothing you type reaches a caller until you publish again. See versions.
8

Talk to it

Scroll to the bottom, to Test call. It is disabled until something is published, because the test panel runs the published version — not the draft in front of you. That is the point: it is the same code path a real caller gets.Press Start call, allow the microphone, and say something. Your browser’s mic connects straight to the agent; no phone number is dialled and no carrier is involved.Things to try in the first thirty seconds:
  • Interrupt the greeting. The agent should stop talking and listen. See turn handling.
  • Ask for the hours, then ask something it cannot know — “how much is a root canal?” — and check that it offers a message instead of inventing a number.
  • Stop talking mid-sentence and wait. That silence is endpointing deciding your turn is over.
Then hang up. A web test call is a real call: it is metered, it is charged the platform fee, it spends your own provider keys, and it is recorded by default. It shows up in the call list like any other.
9

Read the call back

Go to Monitor → Calls and open the call you just had. This is the page you will live in when an agent misbehaves.
  • The transcript, both sides, in order, with every tool call and everything those tools returned.
  • Latency, by stage — how long the speech-to-text took to settle, how long the model took to answer, how long the voice took to start speaking. A slow agent is almost always one of those three, and this tells you which before you start guessing.
  • Cost, split into your provider spend — per provider, per model, from the real tokens, characters and seconds each one metered — and Talqing’s platform fee, which is the only part debited from your credit.
  • The recording, one stereo file with you on one channel and the agent on the other.
  • The exact agent configuration this call ran, frozen. Not the draft you have since edited — which is what makes a week-old call still explainable.
See calls and costs.

​
What to do next

Your agent can talk but it cannot yet do anything — it has no tools, no knowledge and no phone number. Three directions from here, in the order most people take them:

Take a real phone call

Connect your carrier, import a number, point this agent at it, and ring it from your own phone.

Write a better prompt

The prompt is the agent. Voice-specific craft, and how to write tool instructions a model actually follows.

Give it a tool

Look something up, book a slot, transfer to a human, end the call.

Give it knowledge

Crawl your own site so it can answer from your content instead of your prompt.