# Build vs Buy: Open-Source Frameworks, Managed Platforms and Voice APIs

You can build a voice agent on an open-source framework such as Pipecat or LiveKit Agents, assemble it from voice AI APIs, or buy a managed platform. Building gives control and needs engineers. Buying is faster and gives less control.

Updated 2026-10-29. Also on the [web page](https://voiceagents.fyi/guides/build-vs-buy-ai-voice-agents).

There are three ways to get an AI voice agent. Each way moves work between you and a vendor. This guide compares them. Prices change, so check each vendor page.

## The three options

| | Open-source framework | Voice AI APIs | Managed platform |
|---|---|---|---|
| What you get | Code that connects the parts | Individual services: speech, voice, model, telephony | A hosted agent with a dashboard |
| Examples | Pipecat, LiveKit Agents | Deepgram, ElevenLabs, OpenAI Realtime, Twilio | Listings in the voice agent platforms category |
| Who runs it | Your team | Your team, with vendor services | The vendor |
| Control | Full | High | Limited to the platform's options |
| Time to first call | Longer | Medium | Shorter |
| Skills needed | Python or TypeScript engineers, DevOps | Engineers | Operations staff; engineers for integrations |

## Open-source frameworks

### Pipecat

- "An open-source Python framework for building real-time voice and multimodal conversational agents."
- License: BSD-2-Clause.
- Python 3.11 or later.
- Transports include Daily (WebRTC), LiveKit (WebRTC), FastAPI WebSocket and others.
- Telephony serializers: Exotel, Genesys, Plivo, Twilio, Telnyx and Vonage.
- It supports many speech-to-text, LLM, text-to-speech and speech-to-speech services.
- About 16.9k GitHub stars on 2026-10-29.

### LiveKit Agents

- "A framework for building realtime voice AI agents."
- License: Apache-2.0. The LiveKit turn detection models use the separate LiveKit Model License.
- Python, with a separate JS/TS library (AgentsJS).
- A transformer model detects when a user has finished speaking. LiveKit says this "helps to reduce interruptions."
- Works with LiveKit's telephony stack for inbound and outbound phone calls.
- About 15.0k GitHub stars on 2026-10-29.

### What you still pay for

Open-source code has no license fee. You still pay for each service:

- **Hosting.** You can self-host. Or you can use a cloud. LiveKit Cloud lists agent session minutes at $0.01 per minute after the included minutes on its Ship and Scale plans.
- **Speech-to-text.** Deepgram Nova-3 streaming is $0.0077 per minute on pay as you go.
- **Telephony.** Twilio lists US local inbound calls at $0.0085 per minute.
- **Model and voice.** These vary most by your choice.

LiveKit's own phone-call calculator example totals $0.0479 per minute. That example includes session, telephony, LLM, STT, TTS and observability.

## Voice AI APIs

You call each service from your own code. Some APIs cover more than one part of the stack. Examples of public rates on 2026-10-29:

- Deepgram's Voice Agent API lists $0.075 per minute on its Standard tier, pay as you go.
- Twilio lists ConversationRelay at $0.07 per minute. Ask Twilio which parts this rate covers.

APIs remove some of the plumbing. You still build the call logic, the tools and the monitoring.

## Managed platforms

The vendor hosts the agent. You configure it in a dashboard or through an API. Examples of public rates on 2026-10-29:

- Vapi: $0.05 per minute hosting, with models at cost.
- Retell AI: "$0.07-$0.31 / min" for voice agents.
- Bland AI: $0.14 per minute on Start, $0.12 per minute on Build plus $299 a month.

## How to decide

Choose **an open-source framework** when:

- You have engineers who can run real-time audio systems.
- You need to host in your own cloud or region.
- You want to choose every part of the stack.

Choose **voice AI APIs** when:

- You want control of the call logic but not the media servers.
- You already run software that must call the agent.

Choose **a managed platform** when:

- You want a working agent soon.
- You do not have engineers for audio and telephony.
- You prefer one vendor for support and one bill.

## Questions to answer before you build

1. Who is on call when the agent fails at night?
2. Who tests the agent after a model or voice update?
3. Who signs the BAA or data processing agreement with each provider?
4. What is the cost of engineering time per month?
5. What will you do if a provider changes its price or model?

If these answers are unclear, a managed platform is the lower-risk start. You can move to a framework later.

Browse the open-source voice frameworks, voice AI APIs and voice agent platforms categories to compare listings.

## Sources

- [Pipecat, GitHub repository (fetched 2026-10-29)](https://github.com/pipecat-ai/pipecat)
- [LiveKit Agents, GitHub repository (fetched 2026-10-29)](https://github.com/livekit/agents)
- [LiveKit, Pricing (fetched 2026-10-29)](https://livekit.com/pricing)
- [Vapi, Pricing (fetched 2026-10-29)](https://vapi.ai/pricing)
- [Retell AI, Pricing (fetched 2026-10-29)](https://www.retellai.com/pricing)
- [Bland AI, Pricing (fetched 2026-10-29)](https://www.bland.ai/pricing)
- [Deepgram, Pricing (fetched 2026-10-29)](https://deepgram.com/pricing)
- [Twilio, Programmable Voice pricing, United States (fetched 2026-10-29)](https://www.twilio.com/en-us/voice/pricing/us)

## Questions

### Are Pipecat and LiveKit Agents free?

The code is open source. Pipecat uses the BSD-2-Clause license and LiveKit Agents uses Apache-2.0. You still pay for the model, speech, voice, telephony and hosting you use.

### Can an open-source agent answer phone calls?

Yes. Pipecat lists telephony serializers for Twilio, Telnyx, Plivo, Vonage, Exotel and Genesys. LiveKit Agents works with LiveKit's telephony stack to make and receive phone calls.

### Is building always cheaper?

Not always. The per-minute cost can be lower. But your team must run, monitor and update the system. Count engineering time in the total cost.