Skip to main content
Zero Runtime is a runtime, not a framework. With a framework, you write the code and run the servers and GPUs, and manage latency yourself. With a runtime, you write the code and we run the rest. You write the agent; Zero Runtime runs it in real time and handles the hard part. It is serverless for voice AI agents. That puts it between two common options. Self-hosted frameworks give you full control, but you have to run the real-time infrastructure. No-code platforms run that for you, but lock you into a dashboard. Zero Runtime gives you the control of a framework with the ease of a managed platform: real code and your own providers, with nothing to operate.

​
What you get

Real code, not configuration

Agents are ordinary Python, TypeScript, or Go. Your logic and tools run in your own process, so you can use your existing libraries, data, and services.

Your providers, no lock-in

Each pipeline stage is a provider you choose. Swap any provider in a single line.

No real-time infrastructure to operate

Transport, GPUs, turn-taking, and autoscaling run on the runtime. You deploy a worker, not a cluster.

The same model in three languages

Concepts and behavior match across Python, JavaScript, and Go.

​
How it works

A Zero Runtime agent has two halves:
  • The agent is your code: the instructions, logic, and tools you write.
  • The runtime is everything real-time that we run for you: streaming the audio, the GPUs, and the latency of the conversation.
Your code runs in your process. The runtime runs on ours.

​
Building blocks

Every agent uses the same three pieces, named the same way in each SDK.

Agent

Holds your behavior: the instructions, plus on_enter and on_exit hooks that run when a conversation starts and ends.

Pipeline

The voice stack you build from plugins: speech-to-text (STT), an LLM, and text-to-speech (TTS), plus voice activity detection (VAD), turn detection, and noise cancellation.

Invoke

Starts a session on demand. It hosts your agent, connects to the runtime, and handles each conversation.

​
A single turn

A single turn in Zero Runtime: caller audio flows through noise cancellation, voice activity detection, speech-to-text, turn detection, the LLM, and text-to-speech, then back to the caller, who can interrupt at any time.
Each time the caller speaks:
  1. Noise cancellation cleans up the audio.
  2. Voice activity detection (VAD) notices that someone is speaking.
  3. Speech-to-text (STT) turns the speech into text.
  4. Turn detection decides when the caller has finished, so the agent doesn’t cut in too early.
  5. The LLM writes the reply.
  6. Text-to-speech (TTS) speaks it back.
The caller can interrupt at any time. The runtime stops talking and listens again.

​
Who runs what

Who runs what in Zero Runtime: a caller connects to Zero Runtime, which we manage and which runs transport, GPUs, and the speech-to-text, LLM, and text-to-speech pipeline; it connects to your worker, which you run and which holds your agent's instructions, logic, and tools.
Your agent logic and tools run on your side. The hard real-time infrastructure, the part that has to be fast and always available, is ours.

​
When it fits

Zero Runtime is built for phone agents, replacing phone menus (IVR), customer-support voice bots, voice assistants, and automated outbound or inbound calls. It fits when you want to own the agent code but not the real-time infrastructure.