Creating an Agent
Three ways to start a new agent. Pick the easiest for you.
Click Agents → + New Agent. You'll get three options.
Option 1: Pick a template (easiest)
Five pre-built agents you can use as-is or customize:
| Template | Who it pretends to be | When to use |
|---|---|---|
| Maya | Customer support | Refunds, complaints, "where's my order" |
| Arjun | Sales agent | Qualify cold leads, demo bookings |
| Priya | Receptionist | "Hi, please hold while I transfer you" |
| Nisha | Survey caller | NPS, post-purchase feedback |
| Kavi | Appointment booker | Doctor, salon, service-business slots |
Each template has a tested system prompt + good defaults. Pick whichever is closest to your use case — you can change anything later.
Recommended for everyone. Even if it's not exactly right, it's 10 minutes faster than writing from scratch.
Option 2: AI-Generated (medium effort)
You describe what you want; AI writes the prompt.
In the text box, write ONE sentence:
"A friendly Hindi-speaking assistant for a Mumbai dental clinic that books appointments and answers basic questions about cleanings, root canals, and braces."
osmTalk generates:
- Name (you can rename)
- System prompt (review and tweak)
- Welcome message ("Hi, namaste, this is...")
Click Create, then refine.
Use this when your business is specific (you know the exact role) but you don't want to think about prompt-engineering.
Option 3: Blank (most control)
Start with an empty form. You write everything:
- Name — what shows in the dashboard and what the bot calls itself
- System prompt — the long instruction (see Instructions)
- Welcome message — first words the bot says
Use this when you're cloning from an external doc, copying from another vendor, or have a very specific prompt-engineering pattern.
What comes with every new agent (defaults)
When you create an agent, these are auto-set. Change any of them later:
| Default | Why this default | |
|---|---|---|
| AI brain (LLM) | OpenAI gpt-5.4-mini | Fast + cheap + good enough for 95% of calls |
| Ears (STT) | Deepgram nova-3-general | Best English transcription accuracy |
| Voice (TTS) | Deepgram aura-2-helena-en | Natural female English voice |
| Language | English | Switch in Voice tab |
| Recording | Off | Turn on if you need playback / compliance |
| DTMF (keypad) | Off | Turn on if callers will press digits |
| Voicemail detection | Off | Turn on for outbound campaigns |
| Status | Active | Turn off to disable without deleting |
After you create
You'll land on the agent's page. Three things to do, in order:
- Click "Test Call" in the top-right. Talk to your bot. See what it actually does.
- Tweak the system prompt in Configure until it's right. Test again. Repeat 3-4 times — most people get it right by the 3rd try.
- Pick a delivery channel — phone number, website widget, WhatsApp, or campaign.
Picking the right AI brain (LLM)
If you want to change from the default, the trade-offs are:
| Model | Speed | Cost (per 1K tokens) | When to use |
|---|---|---|---|
| OpenAI gpt-5.4-nano | Fastest | ₹0.02 | Simple chats, FAQs |
| OpenAI gpt-5.4-mini ⭐ | Fast | ₹0.04 | Most use cases |
| OpenAI gpt-5.4 | Slower | ₹0.25 | Complex reasoning, agentic tasks |
| Groq llama-3.3-70b | Very fast | ₹0.06 | When you want fast + cheap |
| Anthropic claude-haiku-4-5 | Fast | ₹0.10 | Multilingual, formal tone |
| Anthropic claude-opus-4-7 | Slow | ₹0.50 | Top quality for complex calls |
⭐ = default. Change only if you have a specific reason.
Don't pre-optimize. Start with the default, run 10 calls, then upgrade only if you see issues.
What's next
- Write the system prompt — most important page on this site
- Pick a voice
- Connect a phone number