Modular Backend for Conversational AI
Build agents that call, message and remember in a few lines of code. One SDK between your agent and every customer conversation: telephony, streaming speech, durable journeys, shared context and realtime events. Use one module, or the whole backend.
Every layer between your agent and the customer.
Unpod owns the whole path: carrier-grade telephony, streaming speech, the agent runtime, durable journeys, shared context and observability. End-to-end in under 330ms, one trace across every hop, no vendor chain to stitch. Use one layer, or the whole stack.
Unify customer communication across channels and systems.
A shared context and intelligence layer sits under your agents, journeys and channels, so every call, message and email lands on the same customer record. You build on one backend instead of stitching tools together, and adding a channel never changes your architecture.
Agents
action · voice · chat, spun up per task
Journey Engine
durable flows · waits · branches · signals
Intelligence
definable blocks, anywhere in the pipeline
Calling Engine
numbers · SIP · media · voice router · profiles
Channels
WhatsApp · Email · SMS · Web chat · custom connectors
A WhatsApp message, a 12-minute call and an inbound email are technically different, but they share who participated, which conversation they belong to, what outcome resulted and which event fired. Your application never sees the difference.
Hear an agent handle a real call. Then ship your own on the same rails.
Realtime voice for appointments, reminders and support, on carrier-grade telephony and streaming speech. Tap any card to listen.
# pip install unpod from unpod import Unpod client = Unpod(api_key="up_...") call = client.calls.create( agent_id="agt_123", webhook="https://you.dev/unpod" )
One customer, every channel
Upsert a person once and every identity resolves to them: one entity, one thread, one history across every channel.
Journeys that keep going
Durable flows wait days, branch on outcomes and never call someone who already answered.
Memory that compounds
Outcomes, preferences and history land in one context layer agents read before every turn.
Proven in production
From first playbook to a live phone number.
Data residency
Pin storage to US, EU or India regions. Your choice, per project.
Encrypted end-to-end
TLS in transit, AES-256 at rest. Bring your own KMS keys.
Zero-retention audio
Streams aren’t stored by default. Opt in only if you need recordings.
Your model only
Transcripts train your own model when you ask. Never a shared or third-party one.
See Unpod in Action
Go live in weeks, not quarters. Unpod runs the call, reads the room, and knows exactly when to bring a human in.
The Playbook Writes Itself
A copilot drafts your conversation flow, Talk Live previews it by voice, and every change is versioned. No wall-of-prompt text, no black-box tuning.
- Copilot authors and refines the playbook alongside you
- Talk Live to preview any flow the instant it's ready
Pick a Voice That Fits the Call
Streaming STT and TTS tuned for Indian languages and English. A curated set of voices per business type, previewable in one tap.
- Multiple providers side-by-side: Unpod, ElevenLabs, Cartesia, Sarvam, Smallest, GNANI
- Live latency and quality scores next to every voice, per language
Or Call It Like an API
Same playbook, published as one LLM endpoint the OpenAI SDK can already call: cURL, Python, JavaScript, or Go.
- One endpoint, one model ID, drop-in with any OpenAI-style client
- Stateful sessions via x_session_id, greet-first support built in
One Number. Live Instantly.
The moment your playbook is ready, deploy it as a live voice agent with a real phone number, published and callable in seconds.
- A working phone number the instant you publish. No porting, no waiting
- Or point the OpenAI SDK at it in cURL, Python, JS or Go. One line change
Every Agent, Always Visible
Deployed agents, live sessions, and total minutes at a glance. Track dispatch health and drill into each worker's status in one view.
- Live session counts and dispatch-error trends over 7 days
- Per-agent cards showing active sessions, workers, and deploy age
Every Call, on One Console
Wallet balance, inbound vs outbound, missed calls, spend and success rate. The whole voice stack on one screen, updated live.
- Real-time usage across numbers, agents and languages
- At-a-glance stats: duration, missed calls, success rate, and spend
Runs in your stack.
Point the OpenAI SDK at your trained endpoint, or run the open-source framework in-house. Python, Node, REST or MCP. Text in, text out, any model.
import os
from openai import OpenAI
# Already have an agent loop? Change one line.
client = OpenAI(
base_url="https://api.unpod.dev/v1",
api_key=os.environ["UNPOD_KEY"],
)
reply = client.chat.completions.create(
model="agt_123", # your trained playbook endpoint
messages=[{"role": "user", "content": "Is Friday 4pm free?"}],
)"Yes, Friday 4pm is free. Booked."
Questions before you start.
The backend, the models, integration and security: the things developers ask before shipping.
Place the call. Run the journey. Keep the memory.
Start free in the playground, no login needed. Ship on the whole backend, or export the playbook to your own stack and pay only for metered minutes.