Synthesia puts a talking avatar inside your website’s AI agent

01A face for the agent you already run

Synthesia now lists Interactive Avatars as a generally available API. It brings a responsive video character into a website, product or internal tool, while the developer keeps control of the underlying conversation.

The documented Python integration attaches to a LiveKit voice agent. You supply the language model and speech recognition; Synthesia renders the face, with an optional built-in speech voice. The avatar appears as a participant in the LiveKit room. This is a live conversation rather than a prerecorded presenter reading a fixed script.

02Speak, and the form changes

The official Tools quickstart shows a booking form being filled through conversation. The agent calls a function to update a field, then sends that change to the browser. Typing directly into the form sends the edit back, so the agent can acknowledge the same information.

Both directions share server-side state. That matters when someone says a date aloud, then corrects it with the keyboard: the conversation and interface can refer to one record. The example submits to a log; connecting it to a real booking system or CRM requires additional implementation.

For a product demo, this pattern could let a visitor describe what they need while the page assembles the relevant details. That is an application of the documented tools pattern, not a result from a XANTRANS deployment.

03Give it a source for its answers

A separate RAG example retrieves reference material before composing a spoken answer. It starts with Wikipedia and includes an AWS Bedrock option for a private knowledge base. Updating that reference collection changes what the agent can retrieve without retraining the model.

For a support assistant, the useful content would be the product’s own instructions and policies. Retrieval can make answers more relevant, but Synthesia explicitly describes this grounding as prompt-based rather than a guarantee. A convincing face does not establish that its answer is correct.

04What you need to build it

The official repository provides three starting projects: a minimal conversation, knowledge-base retrieval and interactive tools. The developer overview requires Python 3.10 or later and a Synthesia key with Interactive Avatar access. A direct API route is also documented for developers who manage the session themselves.

There is currently no ready-made iframe to paste into a site. The operational docs specify LiveKit transport, synthetic or personal avatars rather than stock avatars, and bust framing without hand or full-body controls. Mobile is not the primary target.

The examples are local demos, not deployment-ready services. In particular, the Tools example needs authentication around its token endpoint before public use, because starting an agent session incurs costs.

05Check the rate before budgeting

As checked on 23 September, Synthesia’s product page lists $0.12 per minute, while its operational documentation lists $0.10 per minute. Those official figures conflict, so this article does not present either as a confirmed account rate. Verify the price in the workspace before committing a production budget, and include the other services in your chosen agent stack.