A face for your AI, without giving up your stack.
kika.ws is an open layer between your AI stack and a real-time video call, not a black box you rent a personality from.
An open layer, not a bundled product
Any OpenAI-compatible endpoint
In memory, unless recording is turned on
Most avatar products ask you to hand over the whole conversation, their model, their voice, their recognizer, their recording. That is a reasonable product and a difficult dependency.
Every team that deploys kika.ws arrives with constraints we could not have guessed, an existing provider contract, a region requirement, a fine-tuned model that knows their domain, or a policy that a recording must never touch a third party. A bundled stack fails all of those at once.
So we built the thin part instead, the loop that turns streaming audio into a lip-synced face fast enough to hold a conversation, and left every model in it replaceable.
Composable, Not Bundled
The language model, the recognizer and the voice are three replaceable parts. If a provider speaks the OpenAI chat completions shape, it works, and it can be swapped per organization without a redeploy.
A Surface You Can Read
The integration is deliberately small, an endpoint, a key, a model name. A plain API you can read end to end beats a clever one you have to trust.
Latency Is The Product
A talking avatar is judged on how long the silence lasts. We overlap recognition, inference and synthesis so the reply starts before the sentence finishes generating.
Your Data Stays Yours
Media is processed in memory. Recording and transcript retention are off until an organization turns them on, and an on-premise deployment makes no outbound calls at runtime.
- The First Loop
A Lip-Sync Experiment
It started as a test of whether a browser could render a viseme-accurate face fast enough to keep up with a streaming voice.
- The Loop Closes
Under A Second
Speculative transcripts and clause-level synthesis brought the reply under a second, the point it stopped feeling like a demo and started feeling like a conversation.
- Today
Deployments, Not Demos
Organizations, MCP tool calling, Meeting Studio and on-premise deployment shipped, the parts a team asks for once it wants to put an avatar in front of customers.
Working on Something Adjacent?
We read everything you send us. Tell us what you are building.