Introduction
A Go SDK for realtime, multimodal voice AI agents on the server.
agents-go is a Go framework for building realtime, multimodal voice AI agents that run on servers. You give it an agent (instructions + tools) and the models to speak with (STT, LLM, TTS, VAD); it runs the turn loop — listen, think, speak — over a pluggable transport, handling endpointing, interruptions, tool calls, and handoffs between agents.
It's built for large, horizontally-scaled fleets, and it's transport-agnostic: the same agent code runs over a console pipe, a raw WebSocket, a phone call, or a WebRTC room, and behind either its own dispatcher or an external one.
Start here
Get started
Install agents-go and run your first agent in a few lines.
Building agents
The core, one capability per page: sessions, agents & handoffs, tasks, tools, models, turns.
Transports
Get audio in and out — console, WebSocket, WebRTC (roomio), and telephony.
Plugins
The provider & transport catalog — OpenAI, Anthropic, LiveKit, Twilio — and how to add your own.
Deploying
Architecture, workers, scaling, telemetry, and zero-drop drains.
Reference
Exact APIs: Config/AgentSession, Agent, tools, model interfaces, transport, worker.
What you can build
- A voice agent that listens over a WebSocket or WebRTC room, replies with streamed TTS, and is interruptible mid-sentence.
- A telephony agent (e.g. Twilio) that answers calls, collects input, and transfers between specialised agents.
- A multi-agent system where a supervisor hands off to, or delegates a sub-task to, another agent and gets control back.
- A worker fleet that registers with a control plane, reports load, and drains gracefully for zero-drop deploys.