Skip to content

Bananaflow MCP is live.

Read the docs

Pipecat for Voice Agents: Build or Buy?

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Cover for "Pipecat for Voice Agents: Build or Buy?": a white line diagram of build or buy: an open toolbox in isometric with three tools standing in it, a pipe, a plug and a dial, joined by a dashed line through a glass circle holding a question mark to a panel with a phone and a short sound wave, drawn over a soft, out-of-focus pink lotus, dusty rose petals with a blush along the top and soft sky blue at both ends.

Pipecat is an open source Python framework for building voice and multimodal AI agents. It gives you the plumbing that moves audio and text between a phone or browser, a speech-to-text service, a language model and a text-to-speech service. This guide explains how Pipecat works, what a phone agent built on it still needs for Indian calls, how it compares with LiveKit Agents, and when it makes sense to build your own agent or use a platform.

Here is a made-up example. Arjun is the tech lead at a Pune D2C brand that sells kitchenware. He wants a Hindi agent to call cash-on-delivery buyers and confirm each order. On a Saturday he installs Pipecat and talks to a working bot in his browser within the hour. By Monday his list is longer: a phone number, Hindi and Marathi speech that survives a noisy shop, retries for people who did not pick up, and a dashboard his ops team can read. Arjun and his brand are made up for this guide.

This post is about that gap, between a bot that talks in a browser and a phone service you can run every day. We read the Pipecat and LiveKit documentation on 16 November 2026, and every fact about those tools is linked.

What is Pipecat?

Pipecat is an open source framework, written in Python, for real-time voice and multimodal AI agents. Its documentation describes it as an ecosystem that covers creating, deploying and scaling agents that can see, hear and speak, free to use under the BSD-2 license and working with any AI provider and any hosting environment (Pipecat documentation). It is maintained by Daily and the community (Pipecat on GitHub).

The ecosystem has three parts, per the same page:

  • The framework. The Python code that runs on your server and strings the AI services together. The docs say it can orchestrate 150 or more of them.
  • Clients. Software kits for JavaScript, React, React Native, iOS, Android and C++, so a web or mobile app can reach your agent.
  • Pipecat Cloud. A hosted product from Daily for running agents, managed in Daily-hosted regions or inside your own cloud account. The docs also describe self-hosting, so the framework does not depend on it.

Pipecat is a toolkit, not a finished product. You write a Python file, and that file is your agent. For the whole path to a working agent, with or without a framework, see our guide to how to build an AI calling agent.

How does a Pipecat pipeline work?

A Pipecat pipeline is an ordered list of processors, and data moves through it as small packages called frames. The Pipecat guide names four core ideas: frames, frame processors, pipelines and workers (Learning Pipecat).

  • Frames are the data packages: a slice of audio, a piece of text, a chunk of the model's reply. System frames carry control signals.
  • Frame processors do one job each, such as turning audio frames into text frames, or text into speech.
  • Pipelines connect processors in order.
  • Workers run a pipeline. A standalone voice bot is one worker.

The order matters. Audio has to be turned into text before the model can read it, and the reply has to be turned into speech before the caller can hear it. Pipecat's pipeline guide (Pipeline and frame processing) shows this with a short list of steps.

One Pipecat pipeline, in order: transport in, speech to text, the user's words added to the conversation, the language model, text to speech, transport out, and the agent's reply added to the conversation. Frames of audio and text travel between the steps.

Here is the pipeline from Pipecat's own quickstart (Pipecat quickstart), with the comments from the docs. We have not changed any names.

pipeline = Pipeline([
    transport.input(),     # Receive audio from browser
    stt,                   # Speech-to-text (Deepgram)
    user_aggregator,       # Add user message to context
    llm,                   # Language model (OpenAI)
    tts,                   # Text-to-speech (Cartesia)
    transport.output(),    # Send audio back to browser
    assistant_aggregator,  # Add bot response to context
])

Read it top to bottom. The transport hands in the caller's audio, the speech-to-text service writes it as text, and the model replies. The voice turns the reply into audio and the transport sends it back. The two aggregators keep the conversation so far, so the next turn has context. The service names are the quickstart's examples, not a recommendation from us, and the docs say each step can be swapped without rewriting the rest.

How does Pipecat handle turns and interruptions?

It uses voice activity detection plus a choice of turn strategies, and interruptions are on by default. Pipecat's speech input guide says turn starts can come from voice activity detection, from a transcript arriving, or from a minimum number of words. Turn ends can come from a turn detection model (the default, called Smart Turn), from a speech timeout, or from a check that the words so far make a complete thought (Speech input and turn detection).

When a caller speaks over the agent, the docs say the bot stops, in-flight model and voice work is cancelled, and audio that was generated but not yet played is dropped (Interruptions). Those are sensible defaults, not settings tuned to your callers. Our guide to endpointing and turn detection explains why turn checks fail on Indian calls.

What does a minimal Pipecat phone agent need?

It needs five things: a transport, three AI services, a place to run the code, a phone carrier and a way to start calls. Pipecat gives you the first part of that list in the framework. The rest you choose and connect.

  1. A transport. This is how audio gets in and out of the pipeline. Pipecat's docs list transports for Daily, LiveKit, SmallWebRTC, plain WebSockets, a FastAPI WebSocket transport for phone providers, and more (Transports).
  2. Speech to text, a language model and text to speech. Three services, each with its own account and API key. The quickstart asks for three keys for this reason.
  3. A place to run the code. A Pipecat bot is a Python process. The quickstart needs Python 3.11 or later. Running one on your laptop is easy. Running the right number of them on demand is a separate problem, and the docs say so plainly (Deployment overview).
  4. A phone carrier. Pipecat's telephony overview says you can dial in to bots and have them dial out, over both PSTN and SIP, with Daily, Twilio, Telnyx, Plivo and Exotel (Telephony overview).
  5. A trigger. Something has to start or answer each call.

Pipecat's docs describe three ways to connect a phone call:

  • Connection: WebSocket

    How it works
    The carrier streams call audio to your server over a WebSocket
    Providers named in the docs
    Twilio, Telnyx, Plivo, Exotel
    What the docs say about call control
    Basic, managed by the carrier. Best for simple calling and quick prototypes
  • Connection: WebRTC

    How it works
    Peer to peer audio, with Daily's phone service (PSTN) in front
    Providers named in the docs
    Daily
    What the docs say about call control
    Advanced, with track level control. Suits web and mobile clients that also reach phones
  • Connection: SIP

    How it works
    The standard protocol for phone calls over IP
    Providers named in the docs
    Any SIP-capable carrier, with guides that use Daily as the media layer
    What the docs say about call control
    Full control, including transfers and multi-party calls. Suits legacy call centres and warm transfers

Our guide to WebRTC vs WebSocket for voice agents explains why these audio paths differ. For the carrier side, see PRI vs SIP trunk and the SIP trunk glossary entry.

One more fact from the docs matters for planning. A phone bot starts differently from a browser bot. In telephony, the carrier calls your webhook, you answer with carrier-specific instructions, you accept a long-lived WebSocket carrying audio both ways, and you start a bot wired to it. The docs add that callers expect to hear something within a second or two of connecting, so cold starts are a problem and some warm capacity is usually needed (Telephony in production).

How does Pipecat compare with LiveKit Agents?

Both are open source frameworks for building voice agents in code, and they overlap more than they compete. Pipecat is a Python framework built around a pipeline of frame processors. The LiveKit Agents framework lets you add a Python or Node.js program to a LiveKit room as a full realtime participant (LiveKit Agents).

  • Question: What is it?

    Pipecat
    An open source Python framework for voice and multimodal agents, under the BSD-2 license
    LiveKit Agents
    An open source framework for Python and Node.js, part of the LiveKit platform
  • Question: How is an agent built?

    Pipecat
    A pipeline of frame processors: transport, speech to text, model, text to speech
    LiveKit Agents
    A program that joins a LiveKit room as a participant, with plugins for speech to text, language model, text to speech and realtime models
  • Question: Who maintains it?

    Pipecat
    Daily and the community
    LiveKit Agents
    LiveKit
  • Question: Phone calls

    Pipecat
    WebSocket providers (Twilio, Telnyx, Plivo, Exotel), Daily phone service, or SIP
    LiveKit Agents
    SIP telephony that bridges phone calls into rooms, with inbound and outbound trunks and dispatch rules
  • Question: Hosted option

    Pipecat
    Pipecat Cloud, in Daily-hosted regions or your own VPC
    LiveKit Agents
    LiveKit Cloud, or run the agent yourself
  • Question: Testing tools in the docs

    Pipecat
    Scripted and simulated scenarios for evals
    LiveKit Agents
    Unit tests (pytest and Vitest) and LLM-judged simulations
  • Question: Using them together

    Pipecat
    A Pipecat pipeline can carry its audio over LiveKit through the LiveKit transport
    LiveKit Agents
    The agent runs inside a LiveKit room

The choice is not always "Pipecat or LiveKit". Pipecat's own docs include a LiveKit transport that carries a pipeline's audio over LiveKit's WebRTC infrastructure (LiveKit transport). We are not ranking them: both document testing tools, telephony paths and hosted options. Pick the one whose model of an agent fits how your team thinks, and try both on a real call. Read LiveKit's test and evaluate docs and Pipecat's scripted scenarios first, because tests are where teams feel the difference.

What does production add that the framework does not?

A framework gives you the pipeline. A phone service for Indian callers adds a long list on top, and every item is work for your team, even where the framework offers a tool for it. This is our view from how these projects work, not a gap in Pipecat. Both frameworks document testing, telephony and hosting. The point is that someone still has to do each job, and the work is mostly not code in the pipeline.

Two layers. At the bottom, what a framework gives you: a pipeline, transports and service plugins. Above it, the work a phone service still needs: the phone line, Indian-language speech, turn tuning, testing, campaigns, analytics and hosting.
  • Job: The phone line

    What it means for an Indian phone agent
    A carrier account, numbers, a webhook that answers calls, and handling of what the carrier sends. Telephony starts differently from a browser session
  • Job: Indian-language speech

    What it means for an Indian phone agent
    Choosing speech-to-text and voice services that handle Hindi and your other languages on 8 kHz phone audio, and testing them on your own calls. The framework lists 150 or more services, and which one hears your callers best is something only your recordings can tell you
  • Job: Turn taking

    What it means for an Indian phone agent
    Setting waits so a caller reading a policy number is not cut off, and deciding which words count as an interruption. The defaults are a start, not an answer
  • Job: Speed

    What it means for an Indian phone agent
    Measuring from the caller's ear, not the server log, and finding which of the parts is slow
  • Job: Testing

    What it means for an Indian phone agent
    Writing scripted calls in the languages you serve, and running them on every change
  • Job: Campaigns

    What it means for an Indian phone agent
    Lists, calling hours, retries, pacing, and a queue when lines are busy
  • Job: Outcomes and analytics

    What it means for an Indian phone agent
    A result code for every call, extracted fields, a transcript, a recording and a way to send them to your CRM
  • Job: Hosting

    What it means for an Indian phone agent
    A dispatcher, enough warm bots for a morning peak, logs, alerts and upgrades when libraries change

Each row has a guide of ours or a doc page behind it. On the phone line, read the Pipecat telephony overview and our list of cloud telephony providers in India. On language, see our guides to Hindi and Hinglish on calls and voices for Indian languages. Turn taking and speed are covered in endpointing and turn detection and voice agent latency. For testing, see how to test a voice agent. For hosting, Pipecat's deployment overview is the place to start.

None of these is hard alone. The cost is the sum, and the upkeep that follows when a speech service changes its API, a carrier changes its behaviour or a new language is needed. Before you decide, write the list on one page and put your own engineer-weeks against each row.

When should you build and when should you buy?

Build when voice is the product or when you need control that a platform will not give you. Use a platform when voice is a feature of your business and you need calls running soon. These are rules of thumb from how teams usually weigh the cost, so test them against your own situation.

Building is a good fit when:

  • The voice agent is the product you sell, and you need to own every part of it.
  • You need a custom model, a custom voice or a special audio path that no platform exposes.
  • You have engineers who already run real-time systems and want the whole stack in your own cloud.

A platform is a good fit when:

  • Voice is how you reach customers, not what you sell: order confirmations, reminders, enquiries, renewals.
  • You want to start calling within days, not months.
  • Your ops or sales team, not only engineers, will change the script and read the results.

There is a middle path: prototype on a framework to learn what you need, then decide. The first week teaches you the shape of the problem for the price of the services you plug in.

How Bananaflow fits

Bananaflow is an AI voice agent platform for calls in India. We are a platform, not a framework, and this post is about the build-or-buy choice, so here is what a platform covers from the list above, in plain terms.

  • Languages. English (Indian) and 10 Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu, each heard and spoken in its own script. The agent follows the caller's language and can switch mid-call. See our languages page.
  • Your own choices. Each voice agent picks its own speech-to-text, language model and voice, and any one can change without rebuilding the agent.
  • The phone line. Bring your own numbers and carrier, inbound and outbound. See telephony.
  • Campaigns. Bulk calls from a CSV with retries, calling hours, pacing and an A/B split, or calls started from your system by API, n8n or Zapier. See campaigns.
  • Results. Each call ends with one outcome code. You also get a transcript, a recording, extracted fields, a QA score, sentiment and a summary, with webhooks and a BigQuery export. See Call QA.
  • Testing. Try any agent by voice or text in your browser before it takes real calls.

The conversational AI platform page is the overview. Recordings and uploaded files are stored in AWS Mumbai, and call data is not used to train models.

What should you remember about Pipecat?

  • Pipecat is an open source Python framework, free under BSD-2. You pay for the services and hosting you choose.
  • An agent is a pipeline of processors, and frames of audio and text flow through it.
  • A first agent takes an hour. A phone service for Indian callers takes the full list: line, languages, turns, speed, testing, campaigns, outcomes and hosting.
  • Pipecat and LiveKit Agents are both open source and both document telephony and testing. Pipecat can even run over LiveKit.
  • Build when voice is your product. Use a platform when voice is how you reach customers.
  • Whichever you pick, test on your own callers' recordings before you trust any claim, ours included.

FAQ

Frequently asked questions

  • What is Pipecat?

    Pipecat is an open source Python framework for building voice and multimodal AI agents. It moves audio and text between a transport (a browser, an app or a phone call), a speech-to-text service, a language model and a text-to-speech service. It is free to use under the BSD-2 license, and you pay only for the services and hosting you choose. It is maintained by Daily and the community.

  • Is Pipecat free?

    The framework is free and open source under the BSD-2 license, according to its documentation. The cost sits around it: you pay for the speech-to-text, language model and text-to-speech services you connect, for the carrier and phone numbers, and for the servers that run your bots. Pipecat Cloud, a hosted option from Daily, is a separate product.

  • Can Pipecat make and answer phone calls?

    Yes. Pipecat's telephony guide says bots can take inbound calls and place outbound calls over PSTN and SIP, with Daily, Twilio, Telnyx, Plivo and Exotel. Twilio, Telnyx, Plivo and Exotel connect over WebSockets. You still need a carrier account, phone numbers and a way to start and answer each call.

  • What is the difference between Pipecat and LiveKit Agents?

    Both are open source frameworks for voice agents. Pipecat is a Python framework built around a pipeline of frame processors. LiveKit Agents, for Python and Node.js, adds a program to a LiveKit room as a participant. They are not strict rivals, because Pipecat has a LiveKit transport that carries a pipeline's audio over LiveKit. Try both on a real call before you choose.

  • What does a production phone agent need beyond Pipecat?

    A phone line and carrier, speech services tested on your callers' Indian languages, turn timing tuned for pauses and numbers, scripted tests, campaign tools such as retries and calling hours, a record of each call's outcome, and a way to host enough bots at peak. The framework offers tools for some of these, but your team still does the work.

  • Should I build a voice agent with Pipecat or use a platform?

    Build when voice is your product, or when you need control no platform gives you. Use a platform when voice is how you reach customers, as with order confirmations, reminders or enquiries, and you want calls running soon. A common path is to prototype on a framework first, so you learn what you need before you decide.

  • What does a platform like Bananaflow cover that a framework does not?

    Bananaflow is a platform for AI voice calls in India. It covers English (Indian) and 10 Indian languages, bulk campaigns from a CSV with retries and calling hours, an outcome code for every call, transcripts, recordings, a QA score, and testing by voice in the browser. Each voice agent picks its own speech-to-text, language model and voice.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot