Tavus

Most conversational AI companies sell an interface over someone else's model. Tavus describes itself as a research lab and publishes the models it built, which changes what you are evaluating.

The product is a PAL: an agent with a rendered face that holds a real-time conversation, reads what it sees and hears, and decides for itself when to speak and when to wait.

That last capability is the one worth paying attention to, because it is where almost every conversational agent falls apart and where the least marketing attention usually goes.

Tavus: Video Agents That Watch You Back

• PALs — real-time conversational agents with a face, positioned as able to see, hear, act and respond with emotional context.
• Three named models: Phoenix-4.5 for rendering, Raven-1 for perception, Sparrow-2 for turn-taking.
• Turn-taking is the headline, and it is the problem that decides whether a conversational agent feels natural.
• Three ways in: a developer API, a no-code PAL Maker, and fully managed enterprise deployments.
• Founded 2020, describing itself as an AI research lab and claiming 100,000+ developers on the platform.

What Is A PAL And Why Video?

Tavus describes a PAL as a new type of application that can see, hear, act and understand like a coworker or friend, and positions it as whatever the moment calls for — a teammate, an onboarding guide, a customer expert or a companion.

Strip the framing and the technical claim is specific: a rendered human face in a real-time conversation, with a perception model reading the other side of the call.

Video adds two things voice cannot. The first is that the agent can see — an error light, a cracked part, a document held to the camera — which turns a support conversation from description into observation.

The second is presence. People behave differently talking to a face than to a voice, and for roleplay, onboarding and intake that difference is the product rather than decoration.

The cost is equally specific. Video is heavier, harder to render convincingly in real time, and lands in the uncanny valley when it fails — where a voice agent that stumbles merely sounds like a bad phone line, which is why Vocode and other voice-only stacks remain the safer default for pure telephony.

Video lets the agent see the error light rather than ask you to describe it. That is a different product, not a prettier one.

Three Products, Three Very Different Buyers

The Developer API brings together perception, understanding, voice and rendering as building blocks for real-time PAL conversations. This is the route for teams embedding agents inside their own product.

PAL Maker is the no-code path: describe what you want in a sentence, or talk to Charlie — described as the kid inside the machine — and deploy to a website or app in a few clicks.

Enterprise Solutions are bespoke and fully managed, with Tavus designing, building and tuning the agent around a specific workflow and running it in production with integrations and service guarantees.

Those are three genuinely different companies wearing one brand, and the honest read is that they suit different maturity levels rather than different budgets.

The pattern worth noticing is that no-code entry points and managed enterprise delivery usually appear at the same time for a reason: the self-serve product proves the category and the managed one earns the revenue.

Turn-Taking Is The Hard Part Of Voice

Sparrow-2 is described as a real-time conversational understanding model built for natural flow in noisy, multi-speaker environments, modelling semantics, prosody, speaker identity, backchannels, interruptions, background speech, noise and unclear audio to determine when to listen, wait, speak or continue.

That sentence is doing more work than any other on the site, and it deserves unpacking.

Every unbearable conversational agent fails in one of two ways: it interrupts you mid-sentence because you paused for breath, or it leaves a two-second gap after everything you say. Both come from treating silence as the only signal.

Real turn-taking is not a silence threshold. Humans use falling intonation, syntactic completeness, gaze and the difference between a thinking pause and a finished thought — which is why prosody and semantics appear in that model description rather than just voice activity detection.

Backchannels are the detail that gives it away as a serious attempt. Recognising that mm-hm means keep going rather than my turn now is exactly the distinction cheap systems get wrong, and getting it right is most of what makes a conversation feel unremarkable in the good sense.

• Judge a conversational agent on interruptions, not on what it says.
• Test it in noise, with a second person in the room.
• Pause mid-sentence deliberately and see whether it waits.
• Try a backchannel and check the agent does not stop.
• Measure latency to first word, not to full response.

Five Use Cases And Which Actually Fit

Tavus lists five applications: roleplay and L&D agents, HR onboarding, go-to-market agents, customer support and patient intake.

They are not equally strong, and the difference is worth naming.

Roleplay is the best fit on the list. Practice conversations need a patient partner more than a knowledgeable one, mistakes carry no consequence, and a face makes the rehearsal transfer — the same logic behind dedicated tools like Hyperbound for sales practice and Praktika for language.

Intake and onboarding are strong for a different reason: both are structured interviews where the agent asks and the human answers, which is the easiest conversational shape to get right.

Support and go-to-market are the harder cases, because both have an adversarial edge. A frustrated customer or a sceptical buyer will test the agent deliberately, and the escalation path matters more than the conversation quality. That is a deployment design problem rather than a model problem, and the site treats it as neither.

Reading A Hundred Thousand Developer Claim

The homepage states that Tavus powers 100,000+ developers and the world's most forward-thinking companies, which is a number worth reading carefully rather than dismissing.

Developer counts are among the least standardised metrics in software. The figure could mean accounts created, API keys issued, developers who shipped something, or any combination.

That is not a criticism unique to Tavus — practically every developer platform reports the same way — but it does mean the number cannot be compared against a competitor's without knowing both definitions.

What is more informative is the shape of the surrounding evidence. Named models with published research pages, an llms.txt file, a trust centre and separate supplemental terms for a specific product are the artefacts of a company with real deployments rather than a demo.

The claim we would want and do not have is retention: how many of those developers are still building six months later. In a category where the demo is much easier than the deployment, that is the number that would actually settle the question.

Deploying A Video Agent Responsibly

A video agent demo is compelling in a way that skips past the questions a deployment has to answer. That is the risk with this category rather than any weakness in the technology.

This sequence tests the things that decide whether a PAL survives contact with real users.

1. Start a demo conversation and interrupt it deliberately to test turn-taking.
2. Run it with background noise and a second speaker present.
3. Decide how the agent discloses that it is not a person, before you launch.
4. Define the escalation path to a human and test it under frustration.
5. Check the acceptable use policy and supplemental terms against your use case.
6. For patient or HR data, settle privacy and retention through the trust centre first.

Tavus: Products And Scale