bolna

[!NOTE] We are actively looking for maintainers.

Introduction

Bolna is the end-to-end open source production ready framework for quickly building LLM based voice driven conversational applications.

Demo

https://github.com/bolna-ai/bolna/assets/1313096/2237f64f-1c5b-4723-b7e7-d11466e9b226

What is this repository?

This repository contains the entire orchestration platform to build voice AI applications. It technically orchestrates voice conversations using combination of different ASR+LLM+TTS providers and models over websockets.

Components

Bolna helps you create AI Voice Agents which can be instructed to do tasks beginning with:

1. Orchestration platform (this open source repository)
2. Hosted APIs (https://docs.bolna.ai/api-reference/introduction) built on top of this orchestration platform [currently closed source]
3. No-code UI playground at https://platform.bolna.ai/ using the hosted APIs + tailwind CSS [currently closed source]

Development philosophy

1. Any integration, enhancement or feature initially lands on this open source package since it forms the backbone of our Hosted APIs and dashboard
2. Post that we expose APIs or make changes to existing APIs as required for the same
3. Thirdly, we push it to the UI dashboard

graph LR;
    A[Bolna open source] -->B[Hosted APIs];
    B[Hosted APIs] --> C[Hosted Playground]

Supported providers and models

1. Initiating a phone call using telephony providers like Twilio, Plivo, Exotel (coming soon), Vonage (coming soon) etc.
2. Transcribing the conversations using Deepgram, Azure etc.
3. Using LLMs like OpenAI, DeepSeek, Llama, Cohere, Mistral, etc to handle conversations
4. Synthesizing LLM responses back to telephony using AWS Polly, ElevenLabs, Deepgram, OpenAI, Azure, Cartesia, Smallest, Maya, Kalpa etc.

Refer to the docs for a deepdive into all supported providers.

Local example setup [will be moved to a different repository]

A basic local setup includes usage of Twilio or Plivo for telephony. We have dockerized the setup in local_setup/. One will need to populate an environment .env file from .env.sample.

The setup consists of four containers:

1. Telephony web server:
   • Choosing Twilio: for initiating the calls one will need to set up a Twilio account
   • Choosing Plivo: for initiating the calls one will need to set up a Plivo account
2. Bolna server: for creating and handling agents
3. ngrok: for tunneling. One will need to add the authtoken to ngrok-config.yml
4. redis: for persisting agents & prompt data

Quick Start

The easiest way to get started is to use the provided script:

cd local_setup
chmod +x start.sh
./start.sh

This script will check for Docker dependencies, build all services with BuildKit enabled, and start them in detached mode.

Manual Setup

Alternatively, you can manually build and run the services:

1. Make sure you have Docker with Docker Compose V2 installed
2. Enable BuildKit for faster builds:

   export DOCKER_BUILDKIT=1
   export COMPOSE_DOCKER_CLI_BUILD=1
   

3. Build the images:

   docker compose build
   

4. Run the services:

   docker compose up -d
   

To run specific services only:

docker compose up -d bolna-app twilio-app
# or
docker compose up -d bolna-app plivo-app

Once the docker containers are up, you can now start to create your agents and instruct them to initiate calls.

Example agents to create, use and start making calls

You may try out different agents from example.bolna.dev.

Programmatic usage (minimal example)

You can also build and run an agent directly in Python without the local telephony setup.

Example script: examples/simple_assistant.py

import asyncio
from bolna.assistant import Assistant
from bolna.models import (
    Transcriber,
    Synthesizer,
    ElevenLabsConfig,
    LlmAgent,
    SimpleLlmAgent,
)


async def main():
    assistant = Assistant(name="demo_agent")

    # Configure audio input (ASR)
    transcriber = Transcriber(provider="deepgram", model="nova-2", stream=True, language="en")

    # Configure LLM
    llm_agent = LlmAgent(
        agent_type="simple_llm_agent",
        agent_flow_type="streaming",
        llm_config=SimpleLlmAgent(
            provider="openai",
            model="gpt-4o-mini",
            temperature=0.3,
        ),
    )

    # Configure audio output (TTS)
    synthesizer = Synthesizer(
        provider="elevenlabs",
        provider_config=ElevenLabsConfig(voice="George", voice_id="JBFqnCBsd6RMkjVDRZzb", model="eleven_turbo_v2_5"),
        stream=True,
        audio_format="wav",
    )

    # Build a single coherent pipeline: transcriber -> llm -> synthesizer
    assistant.add_task(
        task_type="conversation",
        llm_agent=llm_agent,
        transcriber=transcriber,
        synthesizer=synthesizer,
        enable_textual_input=False,
    )

    # Stream results
    async for chunk in assistant.execute():
        print(chunk)


if __name__ == "__main__":
    asyncio.run(main())

How to run:

export OPENAI_API_KEY=...
export DEEPGRAM_AUTH_TOKEN=...
export ELEVENLABS_API_KEY=...
python examples/simple_assistant.py

This demonstrates orchestration and streaming output. For telephony, use the services in local_setup/.

Note: For REST-based usage (Agent CRUD over HTTP), see API.md in the repo root.

Expected output shape: assistant.execute() is an async generator yielding per-task result dicts (event-like chunks). The exact keys depend on configured tools/providers; treat it as a stream and process incrementally.

Text-only pipeline example

If you want a text-only flow (no transcriber/synthesizer), you can enable a text-only pipeline:

Example script: examples/text_only_assistant.py

import asyncio
from bolna.assistant import Assistant
from bolna.models import LlmAgent, SimpleLlmAgent


async def main():
    assistant = Assistant(name="text_only_agent")

    llm_agent = LlmAgent(
        agent_type="simple_llm_agent",
        agent_flow_type="streaming",
        llm_config=SimpleLlmAgent(
            provider="openai",
            model="gpt-4o-mini",
            temperature=0.2,
        ),
    )

    # No transcriber/synthesizer; enable a text-only pipeline
    assistant.add_task(
        task_type="conversation",
        llm_agent=llm_agent,
        enable_textual_input=True,
    )

    async for chunk in assistant.execute():
        print(chunk)


if __name__ == "__main__":
    asyncio.run(main())

How to run (text-only):

export OPENAI_API_KEY=...
python examples/text_only_assistant.py

Expected output shape: assistant.execute() yields streaming dicts per task step; fields vary by configuration. Handle chunk-by-chunk.

Using your own providers

You can populate the .env file to use your own keys for providers.

Provider	Environment variable to be added in .env file
Deepgram	DEEPGRAM_AUTH_TOKEN
Bolna (self-hosted, OpenAI Realtime transcription protocol)	REALTIME_TRANSCRIBER_URL ({model} is replaced by the model), REALTIME_TRANSCRIBER_KEY

These are the current supported LLM Provider Family: https://github.com/bolna-ai/bolna/blob/10fa26e5985d342eedb5a8985642f12f1cf92a4b/bolna/providers.py#L30-L47

For LiteLLM based LLMs, add either of the following to the .env file depending on your use-case: LITELLM_MODEL_API_KEY: API Key of the LLM LITELLM_MODEL_API_BASE: URL of the hosted LLM LITELLM_MODEL_API_VERSION: API VERSION for LLMs like Azure

For LLMs hosted via VLLM, add the following to the .env file: VLLM_SERVER_BASE_URL: URL of the hosted LLM using VLLM

Provider	Environment variable to be added in .env file
AWS Polly	Accessed from system wide credentials via ~/.aws
Elevenlabs	ELEVENLABS_API_KEY
OpenAI	OPENAI_API_KEY
Deepgram	DEEPGRAM_AUTH_TOKEN
Cartesia	CARTESIA_API_KEY
Smallest	SMALLEST_API_KEY
Maya	MAYA_API_KEY
Kalpa	KALPA_API_KEY
Navana	NAVANA_API_KEY
Custom (the customer's own server, OpenAI speech API)	None: the platform injects the endpoint (synthesizer_base_url) and key (synthesizer_key) per call

Provider	Environment variable to be added in .env file
Twilio	TWILIO_ACCOUNT_SID, TWILIO_AUTH_TOKEN, TWILIO_PHONE_NUMBER
Plivo	PLIVO_AUTH_ID, PLIVO_AUTH_TOKEN, PLIVO_PHONE_NUMBER

Open-source v/s Hosted APIs

We have in the past tried to maintain both the open source and the hosted solution (via APIs and a UI dashboard).

We have fluctuated b/w maintaining this repository purely from a point of time crunch and not interest.

Currently, we are continuing to maintain it for the community and improving the adoption of Voice AI.

Though the repository is completely open source, you can connect with us if interested in managed hosted offerings or more customized solutions.

Extending with other Telephony Providers

In case you wish to extend and add some other Telephony like Vonage, Telnyx, etc. following the guidelines below:

1. Make sure bi-directional streaming is supported by the Telephony provider
2. Add the telephony-specific input handler file in input_handlers/telephony_providers writing custom functions extending from the telephony.py class
   1. This file will mainly contain how different types of event packets are being ingested from the telephony provider
3. Add telephony-specific output handler file in output_handlers/telephony_providers writing custom functions extending from the telephony.py class
   1. This mainly concerns converting audio from the synthesizer class to a supported audio format and streaming it over the websocket provided by the telephony provider
4. Lastly, you'll have to write a dedicated server like the example twilio_api_server.py provided in local_setup to initiate calls over websockets.

Security Acknowledgments

We would like to thank the following individuals for responsibly disclosing security vulnerabilities and helping us keep this project safe:

• @anaskhan54 — June, 2026

Contributing

We love all types of contributions: whether big or small helping in improving this community resource.

1. There are a number of open issues present which can be good ones to start with
2. If you have suggestions for enhancements, wish to contribute a simple fix such as correcting a typo, or want to address an apparent bug, please feel free to initiate a new issue or submit a pull request
3. If you're contemplating a larger change or addition to this repository, be it in terms of its structure or the features, kindly begin by creating a new issue open a new issue :octocat: and outline your proposed changes. This will allow us to engage in a discussion before you dedicate a significant amount of time or effort. Your cooperation and understanding are appreciated