Choosing your approach
Choosing your approach
You now understand the technical differences between LiveKit Agents, Pipecat, and managed platforms like Vapi, Retell, and ElevenLabs Agents. This chapter gives you a practical decision framework for choosing the right approach for your project.
The build vs buy spectrum
The voice AI landscape is not a binary choice between "build everything" and "buy a platform." It is a spectrum:
| Approach | Examples | You own | You outsource |
|---|---|---|---|
| Full DIY | Raw WebRTC + your own orchestration | Everything | Nothing |
| Open-source framework | LiveKit Agents, Pipecat | Agent logic, deployment, infrastructure | Media transport, session orchestration |
| Framework + managed cloud | LiveKit Cloud, Pipecat Cloud | Agent logic | Transport, deployment, scaling |
| Managed platform | Vapi, Retell | Configuration, prompts | Everything else |
| Turnkey solution | ElevenLabs Agents (prompt-based) | Prompt | Everything |
Most teams land somewhere in the middle. The question is where on this spectrum your project belongs.
Decision framework
Work through these five questions in order. Each one narrows the field.
What are you building?
Phone bot or IVR replacement? Managed platforms shine here. Vapi and Retell are built for telephony-first use cases. You can be live in hours with built-in phone number provisioning and call routing.
User-facing product with a frontend? LiveKit has a structural advantage. The integrated client SDKs and Agents UI component library give you automatic state sync between your agent and your UI — something managed platforms do not offer.
Multi-agent or multimodal system? LiveKit's event-driven architecture and multi-agent support handle complex orchestration natively. Managed platforms are typically single-agent, single-modality.
Internal tool or prototype? Any option works. Optimize for speed. A managed platform or even ElevenLabs' prompt-based setup might be fastest.
What is your expected scale?
Volume changes the math dramatically.
Under 1,000 minutes/month: Cost differences are negligible (tens of dollars). Choose based on speed and developer experience, not pricing.
1,000-10,000 minutes/month: Managed platform fees start to add up. A Vapi deployment at $0.28/min costs $2,800/month at 10K minutes. LiveKit Cloud at $0.12/min costs $1,200. The $1,600/month difference may or may not justify the engineering investment.
Over 10,000 minutes/month: The framework approach almost always wins on cost. At 50K minutes, you are saving $5,000-10,000/month. At 100K minutes, the savings fund a full-time engineer.
Over 100,000 minutes/month: Self-hosting LiveKit becomes compelling. You eliminate the platform fee entirely and pay only for compute and AI provider costs.
What does your team look like?
No dedicated engineers: A managed platform is the right call. Retell's visual builder or ElevenLabs' prompt-based setup lets non-technical team members create and iterate.
Small engineering team (1-3 devs): LiveKit Agents or Pipecat with managed cloud deployment. You get framework flexibility without the ops burden.
Dedicated voice/AI engineering team: LiveKit Agents (potentially self-hosted) gives you maximum control. Your team can optimize every layer of the stack.
Mixed team (engineers + non-technical): Consider starting with LiveKit Agents for the core platform and building internal tooling for non-engineers to configure agent behavior (prompts, tools, knowledge bases).
What are your compliance requirements?
Standard SaaS: Any option works. Most platforms offer SOC 2 compliance.
HIPAA/healthcare: LiveKit Cloud, Retell, and Pipecat Cloud all offer HIPAA-compliant configurations. Vapi charges an additional $1,000/month for HIPAA. ElevenLabs' compliance story is less mature.
Data residency requirements: Self-hosting LiveKit is the only option that gives you complete control over where data is processed and stored.
Government/FedRAMP: Self-hosting is likely your only path. No managed voice AI platform currently holds FedRAMP authorization.
What is your timeline?
Need to be live this week: Managed platform. Retell or ElevenLabs can get you from zero to taking calls in hours.
Need to be live this month: LiveKit Agents with LiveKit Cloud. The framework's CLI tools (lk agent create, lk agent deploy) streamline the path to production.
Building for the long term: LiveKit Agents. The upfront investment in learning the framework pays dividends in flexibility, cost savings, and portability over time.
The common path: start managed, migrate to framework
Many successful teams follow a predictable trajectory:
-
Validate with a managed platform. Use Vapi or Retell to prove the use case works. Can your customers actually interact with a voice agent? Is the conversation quality good enough? This takes days, not weeks.
-
Hit platform limits. As you scale or add features, you run into the managed platform's constraints — latency overhead, missing customization options, rising costs, lack of frontend integration.
-
Migrate to a framework. Rebuild on LiveKit Agents with the knowledge you gained from the prototype. You know what your agent needs to do, what edge cases matter, and what your users expect. The rebuild is faster than the initial exploration.
Migration is not wasted work
The code you write on a managed platform is mostly configuration — prompts, tool definitions, conversation flows. These translate directly to LiveKit Agents. Your prompts stay the same. Your tool implementations stay the same. What changes is the orchestration layer, and that is what the framework handles for you.
Migration effort by platform
If you are already on a managed platform and considering moving to LiveKit Agents, here is what the migration looks like:
| From | Effort | Key changes |
|---|---|---|
| Vapi | Medium | Replace Vapi orchestration with AgentSession. Port Squad configurations to multi-agent patterns. Function calling maps directly to LiveKit tools. |
| Retell | Medium | Replace Retell's workflow with Agent classes. Visual flow logic becomes code (more flexible, less visual). |
| ElevenLabs | Low-Medium | Keep ElevenLabs as your TTS provider within LiveKit. Agent logic moves from their config to your code. |
| Pipecat | Low | Both are Python frameworks. Replace pipeline construction with AgentSession declarations. Most provider integrations have LiveKit equivalents. |
Quick reference: which tool for which job
| Scenario | Recommended approach |
|---|---|
| Phone-based customer support, low volume | Retell or Vapi |
| Phone-based support at scale (50K+ min/month) | LiveKit Agents + LiveKit Cloud |
| User-facing product with rich frontend | LiveKit Agents |
| Multilingual voice agent with best-in-class voices | ElevenLabs (or LiveKit + ElevenLabs TTS) |
| Prototype or proof of concept | Whatever is fastest for your team |
| Regulated industry (HIPAA, data residency) | LiveKit Agents (self-hosted or Cloud) |
| Multi-agent orchestration | LiveKit Agents |
| Multimodal (voice + vision + avatars) | LiveKit Agents |
| Low-level pipeline experimentation | Pipecat |
| Visual conversation flow design | Pipecat Flows or Retell |
| No engineering resources | ElevenLabs or Retell |
The bottom line
There is no universally "best" tool. The right choice depends on what you are building, how big it needs to get, who is building it, and how much control you need.
If you are reading this course on LiveKit Academy, you are likely the kind of builder who wants to understand the full stack. LiveKit Agents gives you that understanding while handling the hardest real-time infrastructure problems. It is the choice that scales from prototype to production without requiring a rewrite.
But if you need to prove a concept by Friday and your team has never touched voice AI, start with a managed platform. You can always migrate later — and you will be better informed about what you actually need when you do.
Ready to build?
Head to Course 1.1 (Voice AI Foundations) to build your first LiveKit agent from scratch. You will have a working voice AI application deployed to the cloud by the end of the course.
Test your knowledge
Question 1 of 1
A startup with 2 engineers wants to build a customer-facing voice AI product with a React frontend. They expect to handle 20,000 minutes/month within 6 months. Which approach makes the most sense?