LiveKit Agents

LiveKit Agents is an open-source Python/Node.js framework (github.com/livekit/agents, Apache-2.0, 13,176 stars) for writing realtime voice, video, and multimodal AI agents: you write the orchestration code, and the framework handles the STT-LLM-TTS pipeline, turn detection, interruption handling, and connects everything over LiveKit's own WebRTC transport (github.com/livekit/livekit, also Apache-2.0, 20,530 stars). It's a coding framework, not a no-code bot builder, though LiveKit added a browser-based 'Agent Builder' for prototyping without code in 2026.

Three things stand out. First, real open-source: both the agent framework and the underlying media server are Apache-2.0 and you can self-host the entire stack instead of using LiveKit's cloud, which is not true of most closed voice-agent platforms (Vapi, Bland, Synthflow). Second, the project is unusually active for its size — the 1.7.0 release (2026-08-20) shipped from roughly 30 contributors including several first-time contributors, and the repo was pushed to as recently as 2026-08-26, which rules out the 'abandoned OSS wrapper' pattern common in this space. Third, LiveKit's WebRTC layer has real production pedigree: LiveKit's own technical blog post (Oct 2024) describes in detail how their client/server SDKs carry the audio for ChatGPT's Advanced Voice Mode over LiveKit Cloud, and the framework now ships plugin support (as installable Python extras) for over 70 STT/TTS/LLM/avatar providers, so a buyer isn't locked into one model vendor.

LiveKit Cloud publishes real numbers, not 'contact sales': Build is free (no credit card, 1,000 agent-session minutes, 5 concurrent sessions, 1 free phone number), Ship starts at $50/mo (5,000 minutes included, 20 concurrent sessions, then $0.01/min overage), Scale starts at $500/mo (50,000 minutes, up to 600 concurrent sessions), and Enterprise is custom. Beyond agent minutes, the meter runs on several independent axes — WebRTC participant minutes, downstream data transfer, recording/transcode minutes, and SIP telephony minutes each have their own included allowance and overage rate. A buyer needs to model actual usage across all of these, not just pick a tier off the headline price, because a call-center-scale deployment can hit the concurrent-session or bandwidth caps well before the minutes cap.

LiveKit Inference lets you call STT, TTS, and LLM models from providers like Deepgram, Cartesia, ElevenLabs, AssemblyAI, Google, and OpenAI through one LiveKit API key and one bill, with per-unit prices published on the pricing page (e.g. Deepgram Nova-3 at $0.0042–0.0048/min, ElevenLabs Multilingual v2 TTS at $120–300 per million characters depending on tier). This removes the need to hold separate API keys and accounts with every provider. Whether LiveKit's per-unit price includes a markup over going directly to each provider is not established — we did not independently price every listed model against the provider's own direct rate, so a buyer optimizing hard for AI-inference cost should compare LiveKit Inference's published rate against the provider's own listed rate for their specific model before assuming it's cheaper.

Because the full stack is open source, you can run your own LiveKit server and skip Cloud pricing entirely — self-hosted agents still count against the same 'WebRTC minutes' meter if you connect through LiveKit Cloud for signaling, per the pricing page, but a fully self-hosted deployment (own server plus own agent workers) avoids LiveKit's billing altogether. The tradeoff is operating a Go-based realtime media server (an SFU) yourself, which is a real ops burden that a hosted competitor would otherwise absorb.

LiveKit raised a $100M Series C led by Index Ventures (joined by Salesforce Ventures, Hanabi Capital, Altimeter, and Redpoint Ventures) at a $1 billion valuation, announced 2026-01-22 on LiveKit's own blog. That followed a $45M Series B at a $345M valuation in April 2025. LiveKit states its infrastructure powers Salesforce's Agentforce voice agents and that Tesla uses it for sales, support, and roadside-assistance voice AI — these are vendor claims from LiveKit's own blog that we have not independently corroborated with Salesforce or Tesla. LiveKit also claims livekit-agents is 'downloaded over 1M times a month'; we could not verify this against PyPI's public download-stats API, which returned a rate-limit error during this check.

LiveKit publishes a Vanta-powered Trust Center at trust.livekit.io. The Cloud pricing page states that SOC 2 Type II reports and network penetration test reports are available to Scale and Enterprise customers, and that HIPAA compliance and SSO are available on Scale/Enterprise respectively; these are the vendor's stated compliance posture, not something we independently audited. LiveKit also runs a public status page at status.livekit.io with a visible incident history.