Vocode is an open-source library for building voice-based LLM agents: real-time, back-and-forth spoken conversations with an AI over phone calls, Zoom, or any audio stream. Vocode Core provides the integrations, orchestration and abstractions to assemble a voice application on top of any AI stack, while the companion hosted Vocode API manages agents on live phone calls. The project has around 3,800 stars on GitHub and is MIT licensed.
Vocode chains together speech recognition, an LLM and text-to-speech into a streaming pipeline that handles interruptions, endpointing and turn-taking the way a natural phone conversation requires. It is modular by design: swap in transcribers such as Deepgram or AssemblyAI, language models from OpenAI or Anthropic, and synthesizers such as ElevenLabs, Play.ht, Azure or Cartesia. Telephony is handled through providers like Twilio, so the same agent code can answer a phone number, join a meeting or run in the browser. Python and Node SDKs are available, and the abstractions keep provider-specific quirks out of your application code.
Vocode Core is free and open source under the MIT license. The hosted Vocode API, an enterprise-grade managed service for running agents on phone calls, is a paid product, so the project is listed as Freemium; you also pay the underlying speech and LLM providers you plug in.
Developers building voice assistants, call-centre agents, phone-based booking systems or any product where users talk instead of type. It removes most of the real-time audio plumbing so teams can focus on the agent’s behaviour.