A companion who answers back
One to one, face to face, in real time. She hears what you say, thinks in a model nobody censors, replies in a voice you chose, and her face moves while she says it.
You’re on mic. She’s on camera.
A session is a webcam call with one participant rendered live. Your side captures audio only; her side is generated frame by frame while she speaks.
Wait — can you actually hear me?
Every word. Keep talking — I like your voice.
A depiction of one turn of the production pipeline — the same stages, in the same order, the live session runs.
What happens between you speaking and her answering
Every stage runs on hardware you control. The order below is the order the pipeline actually executes in.
She hears you
Your microphone is transcribed by a self-hosted Whisper model. Nothing goes to a third-party speech service, so the audio of an intimate conversation never leaves infrastructure you control.
faster-whisper · port 8505
She thinks
The transcript goes to a local language model with no content filter and no provider terms to comply with. Character personality, the conversation so far, and memory from earlier sessions are all in the prompt.
Ollama · port 11434
She answers in her own voice
The reply is synthesised through the voice bound to that character — a preset neural voice, or one cloned from a short reference sample. Word timings come back with the audio so the mouth and the sound share one clock.
Piper · GPT-SoVITS
Her face moves while she says it
The audio drives facial keypoints frame by frame, composited against the character portrait and pushed to the browser. Expression, blink and head pose are driven per frame rather than looped.
LivePortrait · port 8504
You can talk over her
A turn is a state machine, not a queue. Interrupt her and the reply she was giving is cut off — she goes back to listening instead of finishing a sentence nobody wants any more.
- Interrupt her mid-sentence. Start talking and the current reply stops; a stale answer is discarded, not played to the end.
- Speak or type. The same pipeline takes a spoken sentence or a typed one — text skips transcription and goes straight to the model.
- Mouth on the clock. Word timings come back with her audio, so lips and sound run off the same clock.
What the companion can do
Memory that survives the session
What she was told last week is in the prompt this week. Conversations persist per character instead of being thrown away when the tab closes.
Per-character memoryNo content filter
The model is uncensored and runs on your hardware. No provider policy layer decides what a user may say, or what she may say back.
Self-hosted weightsOne identity, everywhere
Face, voice and personality are bound to a single character record. The same companion appears in a live session, a generated image and a rendered clip.
Appearance lockDelivered over WebRTC
Sessions are published into a LiveKit room, so a companion can be watched by more than one viewer, each with a separately issued token.
LiveKit · port 7880
Limits worth knowing before you sign up
| Constraint | Where it stands |
|---|---|
| Session length | 10 minutes on Free, 60 minutes on Pro, uncapped on Enterprise. |
| Sessions per month | 100 on Free, 1,000 on Pro, uncapped on Enterprise. |
| Animation engine | LivePortrait drives the face today. The photoreal talking-head engine needs roughly 25 GB of VRAM and is parked until that hardware is in place — sessions fall back automatically rather than failing. |
| Concurrent embedded sessions | Capped per deployment; idle sessions are reclaimed automatically. |
| Full-body animation | Partial. The face pipeline is what is production-ready. |
Asked before signing up
Is she actually uncensored?
Yes. The language model is self-hosted with no content filter and no provider terms layered over it. She does not break character to decline, because nothing in the stack can make her.
How long can a session run?
10 minutes on Free, 60 minutes on Pro, uncapped on Enterprise — with 100, 1,000 and unlimited sessions a month respectively. The full limits are on the pricing page.
Does she remember me between sessions?
Yes. Conversation memory persists per character, so what she was told last week is in the prompt this week.
Where does my audio go?
To a Whisper container on your own infrastructure, and no further. No third-party speech service ever receives the recording — the same engine is available directly as speech to text.
Can she speak in a voice I choose?
Yes — a preset from the voice library, or one cloned from about ten seconds of reference audio. See voice cloning for exactly how that works and what it does not do.
Is her face a video loop?
No. Facial keypoints are generated from the audio she is speaking, frame by frame, at a 30 fps target. The photoreal talking-head engine needs roughly 25 GB of VRAM and is parked until that hardware is in place; sessions fall back to LivePortrait automatically.
Can I put her in my own product?
Yes. Sessions publish into a LiveKit WebRTC room — create one server-side, hand the browser a short-lived token, and she appears on your site instead of ours. Start with the developer overview or the API reference.