One face, whole crew
Talk to one agent. She delegates to specialists, each on its own model with its own permissions.
COMPANION MODE
A 3D AI companion that runs your crew.
Talk to one face. A crew does the work. Give any agent a 3D face and a voice. You talk to her; she gets your whole crew moving and tells you how it went, out loud.
She breathes, blinks and looks around while you think. She listens while you talk. She looks thoughtful while the agents work. Then she answers in her own voice, her mouth shaping the words, a wave or a shrug landing right on cue. And she is not a talking picture on top of a chatbot: she's a full agent, with tools, files, your servers and a crew behind her.
Belle in the console: listening, thinking, speaking
NOT A CHATBOT SKIN
She doesn't just chat. She gets it done.
Most talking avatars sit on top of a chat model. Yours sits on top of a real agent.
Ask her to research something, fix a file on your server, or put the crew on it. She uses the same tools as any agent: web search, files, your servers, schedules, plugins, and her colleagues. She hands the job to the right specialist in the same turn, waits for the report, then tells you the result in a few spoken sentences. Anything long, like a report or some code, lands in a file or the chat instead of in your ear.
Same agent, two tabs
Flip between Companion and Chat. It's one conversation: spoken on one side, clean text on the other.
Your agent, your rules
Put a face on an agent you already have, or start fresh. Her personality is the agent's own.
- you (voice)Belle, the nightly backup failed. Can you find out why and get it fixed?
- Belle (voice)"Hmm, let me look."
- BelleHands kestrel the job: "Nightly backup on db-01 failed. Find the cause, fix it, report back."
- systemBelle is thinking while kestrel works
- kestrelReads the backup log on your server
- kestrelFixes the backup settings
- kestrelRuns a test backup
- kestrelReports back: "The backup disk was full. Old copies cleared; the test backup passes."
- Belle (voice)[happy] "Found it. The backup target was full. Kestrel cleared out old copies and a test run passes now. The details are in the chat."
Belle hands the job to kestrel, the specialist allowed on your servers. Her [happy] cue is performed in her voice; it never shows in the text.
THE CAST
Pick a 3D avatar, or bring your own.
Four characters ready to go, each with her own look, moves, personality and a suggested voice.
Belle
Belle
Warm, playful, cheers you on.
Rose
Rose
Calm, poised, helps you think.
Seed-san
Seed-san
Bright, curious, loves how things work.
Polydancer
Polydancer
Lively, confident, keeps you moving.
Click Talk and she becomes a new agent, named whatever you like. Or click Talk as… to give an agent you already have her face. If you've connected her suggested voice vendor, she speaks with that voice straight away.
Upload your own VRM avatar.
Upload a VRM from VRoid, Blender or UniVRM, or any humanoid GLB. We map the skeleton for you, with an editor if you want to fix a bone by hand. Add Mixamo FBX clips for new moves.
- Tails, ears and wings get their own animations, and sway gently on their own when nothing else moves them.
- VRM avatars get the full treatment: facial expressions, lip-sync and hair and clothing physics.
Room for 50 companions per account, and avatars up to 250 MB.
THE PERFORMANCE
She never freezes.
Not a picture that moves its lips. A performance, from the first blink to the last word.
While you talk, she listens and tilts her head. While the agent works, she looks thoughtful, and after a few seconds of real work she settles into a working stance. When she answers, her mouth shapes the sounds, her brows lift on the stressed words, a gesture lands on the right line, and she laughs where the agent wrote a laugh. Her eyes come back to you just before she finishes, the way a person's do.
Grounded in how people move
Blinks, glances, nods and pauses follow the rhythm of real people.
Gestures that fit her
A gesture is picked for each sentence, steered by her personality. It's picked while her first sentence is already playing, so her voice never waits for her body.
In sync on any headphones
Mouth and gestures follow the sound you actually hear, so they stay in step even on Bluetooth.
Emotion in the voice
Happy, curious, whispering, a sigh, a gasp, a real laugh: performed out loud, never printed in the text.
Her moves, your way
Four motion styles to start from: Playful, Confident, Professional, Energetic. Copy one, swap clips, or build a set from your own avatar's animations.
HAIL HER
Talk to your AI agent out loud.
Speak to her like a person. She waits until you've actually finished.
With Deepgram connected, your words appear on screen as you say them, and she knows when you're done talking, so a pause mid-sentence doesn't cut you off. Your audio goes straight from your browser to Deepgram; it never passes through our servers. With another speech-to-text provider she transcribes each phrase the moment you finish it.
- Stop her mid-sentence. The agent's history records exactly how much you heard, so she never assumes you caught the rest.
- Talk over her. Flip one switch when you're wearing headphones and your voice alone interrupts her.
- Natural pauses. On voice turns she'll often fill a wait with a quick 'Hmm, let me see' while the agent thinks.
- Type any time. Same conversation, keyboard or voice, your choice per message.
VOICES
Your voice vendor. Your key. Your choice.
Pick who speaks and who listens, for the whole account or per companion.
| Vendor | Speaks | Listens | Words appear as you talk |
|---|---|---|---|
| ElevenLabs | Yes | — | — |
| fish.audio | Yes | Yes | — |
| Deepgram | Yes | Yes | Yes |
| OpenAI-compatible (e.g. Whisper) | Yes | Yes | — |
| Any voice plugin | Yes | Yes | — |
Browse and preview each vendor's voices right in the console, with search and samples. Use one vendor to speak and another to listen. A character suggests a voice; you can always pick another.
A companion pack is data, not code: a character with her look, moves, personality and voice suggestions, an avatar, or a motion set. Install it and it shows up in your gallery. Motion sets stay live: when the author improves them, every companion using them moves the new way. Built a motion set you love? One button exports it as a ready-to-publish plugin. Voice vendors can be plugins as well.
BUILT IN
A 3D companion, in the box.
Other agents can talk. OpenClaw has Talk Mode, and Hermes Agent speaks on Telegram and Discord. Neither ships a 3D character of its own; OpenClaw's avatars are community add-ons. Here the face, the voice, the lip-sync and the gestures come with the product, and the character is the full agent.
Give your crew a face.
Sign up, pick Belle, Rose, Seed-san or Polydancer, and say hello. Connect a voice vendor and she answers out loud.