rtp-agent

rtp-agent is a Go runtime for building realtime AI agents. It follows the useful runtime behavior of LiveKit Agents where that behavior is implemented and tested in this repository: agent/session orchestration, streaming LLM/STT/TTS boundaries, tools, interruptions, worker lifecycle, room I/O, telemetry, and provider adapters.

This documentation describes the Go source API at tag v0.0.67. The tag exists in this checkout. The documentation branch may contain docs-only commits after that tag; the diff from v0.0.67 to this documentation update contains no Go source, go.mod, or go.sum changes.

What is implemented

The public composition path is:

1. Build an app.AppConfig, usually from app.DefaultConfigFromEnv().
2. Call app.Init or app.NewApp.
3. Run the resulting worker.AgentServer through interface/cli.RunApp.

For lower-level composition, the source exposes agent.NewAgent, agent.NewAgentSession, model interfaces in core/llm, core/stt, and core/tts, and provider constructors in adapter/<provider>.

Source of truth

The Go source is the API contract for this version. These docs intentionally avoid conceptual APIs that are not present in source, including NewProvider, agent.NewSession, agent.WithLLM, agent.WithSTT, and agent.WithTTS.

Behavioral parity with LiveKit Agents is a project goal, not a blanket compatibility claim. When parity-sensitive behavior is documented, it should be backed by source, tests, or a row in scripts/parity-fixtures/test-cases.tsv.

Documentation map

• Tutorials: start with the basic agent and run it locally.
• How-to guides: configure models, tools, workers, and runtime behavior.
• Explanation: understand architecture, lifecycle, provider boundaries, and parity.
• Reference: inspect packages, constructors, environment variables, and provider capabilities.