Vibecode avoice agentin minutes

Name: Gradbot
Author: Gradium

50 lines of code. Any OpenAI-compatible LLM. VAD, turn-taking, fillers, interruptions, and low latency handled for you.

Fork Me See what people build with it

Live demos

Prototype
with Gradbot

Each demo below runs a live voice AI agent built with Gradbot. Click to start, then talk to them.

Talk to NPCs in a 3D world.
Reserve rooms with voice.
Record a short message, upload a photo, and get a love letter in your cloned voice.

See the app gallery

Click to launch demo

Features

Ready when you fork it

Everything’s already here, so you don’t spend weeks building it yourself.

STT, LLM inference, and TTS coordinated in real time by a Rust core.

It knows when to speak and when to listen. Interrupt it mid-sentence and it stops to hear you out.

When your agent looks something up or makes a booking, it keeps the conversation going instead of going silent.

Plug in the services you already use, and your agent can use them during the conversation.

Run the agent on a server and connect clients, phone lines or browsers to it over the network.

Bring the model you already use: any endpoint that speaks the OpenAI API works as the agent's brain.

Features

Ready when
you fork it

Everything’s already here, so you don’t spend weeks building it yourself.

Bring your
code agent

Point your coding agent at the Gradbot skill, it carries the best practices and gives you the event loop for voice agents. You write the logic, it handles the rest.

Get the skill

Add a Gradbot voice agent to this project.

Read the Gradbot skill at
https://github.com/gradium-ai/gradbot
first and follow it over anything you assume.
Install gradbot with pip and read the Gradium API key
from an environment variable, never hardcode it.

Then implement the agent, in this order:

1. Pick a flagship voice with gradbot.flagship_voice
   and build a SessionConfig from it: its voice_id,
   its language, and the instructions below.
2. Start the event loop with gradbot.run, OggOpus in
   and out, and keep the input and output handles it
   returns.
3. Stream microphone audio into the input handle and
   play what comes out of the output handle as it
   arrives. Do not buffer a full reply before playing.
4. Leave turn-taking, interruptions and tool calls to
   Gradbot. Put the product logic in the instructions
   and in tools, nowhere else.

Instructions for the agent:
"You are a helpful assistant."

Keep it small: one module, typed, with clear errors
when the key is missing or the session drops.
Do not add a UI unless I ask for one.

import gradbot

voice = gradbot.flagship_voice("Harper")
config = gradbot.SessionConfig(
    voice_id=voice.voice_id,
    instructions="You are a helpful assistant.",
    language=voice.language,
)

input_handle, output_handle = await gradbot.run(
    session_config=config,
    input_format=gradbot.AudioFormat.OggOpus,
    output_format=gradbot.AudioFormat.OggOpus,
)

Preview

Harper

Community

Don’t
build alone

Gradbot is open source and shaped by the people using it.

Join the Discord Explore the community

Apps built
by the community

See all apps Submit your app

Broskis

Photograph everyday objects to turn them into weapons in a two-player arena game. Gradium Text-to-Speech and Voice Design provide the match commentary.

Try it