Tavus

Tavus builds the whole video agent stack itself. Instead of putting a talking avatar on top of a chatbot, it trains its own foundational models (the site lists Phoenix-4.5, Raven-1, Sparrow-2 and a preview of Griffin) and splits each agent, which it calls a PAL, into parts. There's a Face that reacts and emotes in real time, a Voice with human rhythm and inflection, Eyes that read your expressions and surroundings, and Ears that pick up tone and hesitation and can tell "the difference between a pause and a finished thought." On top of that sits a personality layer with rules, memory across conversations, a knowledge base you upload, and either Tavus's hosted LLMs or one you bring yourself.

That mechanic is the point. Most avatar tools render a video from a script. Tavus is built for live, two-way conversation, and its headline claim reflects that: 48% of people in its test thought Griffin was a real person, which the company calls the first pass of a real-time video Turing test. The agent can also do things during the call. It presents slide decks by itself, generates charts, quizzes and scheduling pages on screen with Magic Canvas, opens your product in a browser and clicks through it, and joins Zoom, Google Meet or Teams calls as an on-camera participant.

You can get in three ways. Developers use the API, which Tavus says is used by 150,000+ developers. Non-technical teams can use PAL Maker, a no-code builder with a guide named Charlie that walks you through it. Larger companies can book a fully managed Enterprise build. The listed use cases are specific: AI SDRs, healthcare agents, interview agents and L&D agents. The free consumer tier gives you a feel for the experience, but at 15 minutes of voice and video it's a demo, not a working allowance. If you're weighing Tavus for a product, the real question is whether the developer API fits your latency and cost budget. The pages we reviewed don't settle that.