Pewee

Conversational voice for AI agents

Self-hosted. Sub-200ms latency. Zero API dependencies.

Join the Waitlist

The problem

Voice for agents is broken

APIs deprecate

Gemini Live deprecated March 2026. Your agent went mute.

Latency kills

700ms pipelines feel robotic. Users notice.

Tool calling fails

85% accuracy in voice mode. Not good enough.

Architecture

One WebSocket. Full conversation.

Three components, one pipeline, zero external calls.

Ears

faster-whisper streaming STT with speculative processing

Brain

Qwen3.5 + Speculative Engine starts responding while user still speaks

Mouth

Chatterbox Turbo — beats ElevenLabs in blind tests (63.75% preference)

Performance

Numbers that matter

<200ms

average latency
(speculative hit)

63.75%

preferred over ElevenLabs
in blind tests

<$0.01

per minute
all-in

external API
dependencies

Developer experience

Your agent speaks in 10 lines

Connect, stream, done.

agent.py

import pewee
 
agent = pewee.Agent(
    api_key="pk_...",
    personality="You are a helpful desktop assistant.",
    voice="es-mx-jorge",
    tools=[create_task, delete_task, list_tasks],
)
 
session = await agent.connect()
 
async for chunk in microphone.stream():
    session.send_audio(chunk)

Early access

Get early access

We'll reach out when we're ready. No spam.

Thanks! We'll be in touch soon.