LiveKit Agents

Executive Summary

LiveKit Agents is an open-source framework for building realtime voice, video, and physical AI agents. Unlike managed platforms like Vapi or ElevenLabs, LiveKit gives developers full control over provider selection and deployment. The company raised a $100M Series C at a $1 billion valuation in January 2026 led by Index Ventures, bringing total funding to $122.5M+, and counts OpenAI (ChatGPT voice mode), xAI, Salesforce, and Tesla among its customers.^[1][2] The framework ships frequently — the 1.5.x line reached v1.5.17 on June 3, 2026 — and the repository has grown to roughly 10.9K GitHub stars and 3.2K forks as of June 2026.^[3]

---

Product Overview

LiveKit started as realtime video/audio infrastructure and expanded into AI agents with the Agents framework. The framework allows developers to build AI-driven applications that can see (video), hear (audio), and speak (TTS) in realtime, with full flexibility in choosing AI providers.^[4]

The platform supports both managed cloud deployment (LiveKit Cloud) and self-hosting, making it attractive to teams wanting control without building infrastructure from scratch. Since early 2026, LiveKit has layered a managed Agent Platform on top of the framework: LiveKit Inference provides bundled STT/LLM/TTS model access under a single API and bill, and Agent Builder lets teams prototype and deploy agents from the browser without writing code.^[5][6]

Key Capabilities

Framework Components

---

Technical Architecture

LiveKit Agents runs as a server-side process that connects to LiveKit's realtime infrastructure (cloud or self-hosted). Agents can use any combination of AI providers through plugins.

┌─────────────────────────────────────────────────┐
│              Client Applications                 │
│       Web | Mobile | Phone (SIP)                │
├─────────────────────────────────────────────────┤
│         LiveKit Infrastructure                   │
│    (Cloud or Self-Hosted)                       │
│  ┌───────────────────────────────────────────┐  │
│  │         Realtime Media Routing            │  │
│  │    Audio/Video Streams ↔ Agents           │  │
│  └───────────────────────────────────────────┘  │
├─────────────────────────────────────────────────┤
│              LiveKit Agents                      │
│  ┌─────────────────────────────────────────────┐│
│  │         Agent Process                       ││
│  │  ┌─────────┐ ┌─────────┐ ┌─────────┐       ││
│  │  │   STT   │ │   LLM   │ │   TTS   │       ││
│  │  │(plugin) │ │(plugin) │ │(plugin) │       ││
│  │  └─────────┘ └─────────┘ └─────────┘       ││
│  │        ↓           ↓           ↓           ││
│  │  Deepgram    OpenAI      ElevenLabs        ││
│  │  AssemblyAI  Anthropic   Cartesia          ││
│  │  Whisper     Groq        Rime              ││
│  └─────────────────────────────────────────────┘│
└─────────────────────────────────────────────────┘

Deployment Options

---

Strengths

• Open source — Apache 2.0 license; no vendor lock-in; self-hostable^[3]
• Provider flexibility — Mix any STT, LLM, TTS providers through plugins, or use bundled LiveKit Inference
• Strong funding — $100M Series C at a $1B valuation from Index Ventures and Salesforce Ventures^[1]
• Marquee customers — OpenAI's ChatGPT voice mode, xAI, Salesforce, and Tesla run on LiveKit^[1]
• Multi-modal — Voice, video, and physical agents in one framework
• Multi-agent — Built-in support for agent handoff and specialization
• Active community — 10.9K+ GitHub stars, 3.2K forks, active Discord^[3]

---

Cautions

• More complexity — Requires more setup than managed platforms like Vapi; third-party comparisons describe LiveKit as "hands you a communication room and lets your agent sit inside it" versus Vapi's finished product^[7]
• DIY responsibility — You manage provider accounts, rate limits, failover (unless using LiveKit Inference)
• Cost questions at scale — Skeptics have argued the managed layer is expensive for high-volume workloads^[8]
• Learning curve — Framework concepts require developer investment
• Cloud costs stack — LiveKit + STT + LLM + TTS costs add up
• No-code is new — Agent Builder is recent; the platform remains primarily developer-focused

---

Pricing & Licensing

Open Source (Self-Hosted):

• Framework: Free (Apache 2.0)
• Infrastructure: Your costs

LiveKit Cloud (as of June 2026):^[6]

Model costs via LiveKit Inference (examples):^[6]

• STT: ~$0.0058/minute (Deepgram Nova-3)
• LLM: ~$0.0203/minute (OpenAI ChatGPT Latest); Anthropic models also available via plugin
• TTS: ~$0.03/minute (Cartesia Sonic)

Typical total: roughly $0.05-0.10/minute depending on providers — the January 2026 pricing overhaul replaced the old $0.004/min audio metering with plan-based agent session pricing.

---

What Developers Say

From the Show HN thread for the framework (the stack behind ChatGPT's Advanced Voice):^[8]

Praise:

"You guys are honestly the best" — shayps, Hacker News

LiveKit CEO Russ d'Sa, responding to "why not just build it yourself" skepticism in the same thread:

"There's a ton of complexity under the 'relatively simple use case' when you get to a global, 200M+ user scale." — russ, Hacker News

Criticism:

"I'd expect OpenAI to migrate off of LiveKit within 6 months. LiveKit is too expensive." — throw14082020, Hacker News

(That prediction did not hold — OpenAI remained a flagship customer through the January 2026 Series C.^[1])

"80% of the times I'm experiencing choppy audio on my iPhone 15 Pro Max... I wonder if this is because bugs or the crazy load livekit may be going through." — gastonmorixe, Hacker News

Third-party evaluations position LiveKit as the control-maximalist option: AssemblyAI's 2026 comparison concludes LiveKit "gives you the most infrastructure-level control, especially if you self-host," while recommending Vapi for teams that "would rather configure than code."^[7]

---

Competitive Positioning

Direct Competitors

When to Choose LiveKit Agents

• Choose LiveKit when: You want open-source control, provider flexibility, or need self-hosting
• Choose Vapi when: You want managed platform with less setup
• Choose Retell when: You want phone-focused with no-code option
• Choose ElevenLabs when: Voice quality is the top priority

---

Ideal Customer Profile

Best fit:

• Developer teams wanting full control over voice AI stack
• Organizations requiring self-hosted deployment
• Multi-modal applications (voice + video agents)
• Teams wanting to mix best-of-breed providers
• Companies avoiding vendor lock-in

Poor fit:

• Non-technical teams needing no-code solutions (though Agent Builder is narrowing this gap)
• Teams wanting fastest path to production
• Simple phone automation use cases
• Organizations preferring single-vendor simplicity

---

Viability Assessment

LiveKit's combination of a $1B valuation, open-source model, and validation from OpenAI positions it well for long-term success in the voice AI infrastructure space. The 2026 expansion into Agent Platform, Inference, and Agent Builder signals a move up-stack from infrastructure toward the managed platforms it competes with.

---

Bottom Line

LiveKit Agents is the best choice for developer teams wanting maximum control over their voice AI stack. The open-source framework with provider flexibility means no vendor lock-in and the ability to mix best-of-breed components. Self-hosting options make it suitable for organizations with data sovereignty requirements, and the new LiveKit Inference and Agent Builder layers reduce the setup tax that used to be the main objection.

The trade-off is still complexity—more setup than managed platforms, more responsibility for provider management, and a steeper learning curve. Costs also stack (plan fee + agent session minutes + model minutes), so model the per-minute economics before committing at scale. For teams with developer resources who want control, it's excellent. For teams wanting the fastest path to production, managed alternatives like Vapi or Retell may be more practical.

Recommended for: Developer teams wanting open-source voice AI with provider flexibility, self-hosting options, and multi-modal capabilities.

Not recommended for: Non-technical teams, those wanting fastest setup, or simple phone-only use cases.

Outlook: Very positive — $1B valuation, marquee customers (OpenAI, xAI, Tesla), and a fast-shipping open-source core; watch whether the up-stack push into managed agent platforms dilutes its neutral-infrastructure positioning.

---

Research by Ry Walker Research