A face thatlistens back.

A live video call between the visitor's browser and an AI participant. The video is generated frame by frame in Tavus's cloud and streamed over WebRTC (the Daily video network), so it works in any modern browser with no app or plugin.

No. Lips, eyes and expressions are rendered live from what she is saying and hearing. Only the face's appearance is learned in advance, from a short training video or a photo.

Tavus doesn't publish a number. We measured 2 to 3 seconds from the end of your sentence to the start of her reply, and 6 to 9 seconds for the first frame when a call starts. The start delay can be hidden by preparing the call while the page loads.

Yes. The avatar can call any OpenAI-compatible endpoint that streams responses (/chat/completions over SSE). Set model, base_url and api_key on the LLM layer. The endpoint doesn't even have to be a single model: it can be Bedu's full agent behind an OpenAI-shaped API. When perception is on, Tavus adds the visitor's emotions and appearance to the messages it sends.

Source: Tavus docs, LLM layer

Not directly, and it shouldn't. Tavus never needs database credentials. Two clean routes: Bedu's LLM endpoint does the lookup itself, or a tool sends a signed HTTPS request to a Bedu endpoint that queries the database and returns the answer.

Yes. Tavus has a knowledge base: PDF, Word, Excel, CSV, PowerPoint, images, or a website it crawls (up to 100 pages). Documents are attached per call, and retrieval can be tuned for speed or quality. The catch for us: the knowledge base is English-only for now, so Arabic deployments should do retrieval in Bedu's LLM layer.

Source: Tavus docs, knowledge base

Yes. Every call can carry its own context and greeting, for example the guest's name, room and booking dates. During the call, the page can add context or inject a line for her to respond to.

Yes, if we tag the visitor with a stable id. Tavus builds a memory per visitor and per avatar after each call; it can be deleted, but not edited. Pinned facts can be added by hand.

Yes, through tools. A tool is defined with a name, a description and a JSON schema. When she decides to use it, Tavus either sends the call to the web page or calls our HTTPS endpoint directly (with HMAC signing). She can say a filler line like “let me check that” while it runs, then answer from the result.

Source: Tavus docs, tools

Ending the call is a built-in tool. A handover to staff would be one of our own tools, for example creating a callback request in the CRM or opening a WhatsApp chat.

Yes. The page can make her respond to text as if the visitor said it, make her say an exact line, interrupt her, or update her context. This is how a page could react to clicks, such as “the guest just opened the spa page”.

Yes, if the visitor allows the camera. The perception model reads expressions, appearance and a shared screen, and can trigger tools on what it sees, for example when someone holds up a document.

It depends on the plan: 1 call at a time on Free and Starter, 3 on Builder, 10 on Growth, 15 on Business, and custom on Enterprise. A public government service needs Enterprise.

5 minutes on Free, 10 on Starter, 15 on Builder, up to 60 on Growth and above.

Conversation minutes. Plans include a monthly allowance, then $0.26 to $0.35 a minute depending on the plan. A custom LLM, knowledge base and tools don't add Tavus charges, but Bedu's own LLM costs apply.

Source: tavus.io/pricing, 18 Oct 2026. Tavus's page shows two price grids; confirm at checkout.

Tavus's trust centre lists SOC 2 Type 2, HIPAA and GDPR compliance, and they offer a BAA.

Source: trust.tavus.io

Tavus says it doesn't train on customer data. Conversations can be hard-deleted through the API. Zero Data Retention, where nothing is stored, is an account setting their team switches on, typically for enterprise customers.

Only if we switch it on, and then the recording goes into our own storage bucket (AWS, Google Cloud or Azure), not Tavus's.

Not documented. As far as we know it runs on US cloud, and Tavus documents no regional or private deployment option. For government buyers this is the first question to put to Tavus Enterprise, and a possible deal-breaker.

Her instructions require it when asked. Tavus also has built-in disclosure settings (spoken and visual) and an EU AI Act mode.

For Phoenix-4.5: one continuous chest-up shot of about 30 seconds speaking and 30 seconds still, at least 1080p and 25 fps. A face can also be made from a single photo and is usable within minutes, then refines over a few hours. We confirm we hold the rights to the likeness.

Source: Tavus docs, faces

Yes. Tavus can even generate a face from a text prompt, and realistic as well as stylised humans are allowed. Animals, mascots and other non-human characters are not, and faces that look under 18 are rejected.

42 spoken languages, including Arabic and Hebrew. Dialect is not something Tavus documents: we get natural Emirati Arabic by instructing the brain to speak khaleeji, and it worked well in our tests. Kazakh isn't covered by the default voice engine.

Yes: a Tavus voice, a cloned voice, or a voice from ElevenLabs or Cartesia, including private voices on the client's own account.

The real-time engine is Tavus, and we rent it. MEWE owns the persona design, the custom face or character, the voice direction, the knowledge structure, the testing harness and the front end. Bedu owns the agent layer: its LLM, tools and client integrations.

Not in our setup. The face is a swappable layer behind our own page and backend, so we can move to Griffin, or to another engine, without rebuilding the experience.