parlot-instrumentation-livekit

MIT OpenTelemetry instrumentation for LiveKit Agents — the primary path into Parlot, the observability platform for production voice AI.

Know if your voice agent is actually working: conversational turns (not flat traces), multi-agent handoff graph, STT / LLM / TTS waterfall, optional call recording, and goal completion with an evidence trail to the proving turns and audio.

Works with LiveKit Cloud, fully self-hosted media, and LiveKit on Telnyx.

pip install parlot-instrumentation-livekit
# or: uv add parlot-instrumentation-livekit
# or: pip install "parlot[livekit]"

from parlot.instrumentation.livekit import configure

configure()

Docs: LiveKit guide · Concepts · Quick Start · parlot.ai

Integration checklist

Two steps are required for full Parlot behavior. All SDK examples follow this pattern.

1. configure() at import (required)

Call before constructing AgentSession — patches AgentSession.__init__, sets up OTLP export to PARLOT_ENDPOINT, and installs turn/handoff/close event hooks.

from parlot.instrumentation.livekit import configure

configure()

Optional: agent_id= for canonical deployment identity and version= for deployment version. Full resolution order is in the SDK README.

What to expect for version

configure(version=...) resolves once at startup (first non-empty wins): version= kwarg → __main__.__version__ / VERSION → PARLOT_AGENT_VERSION → local git short SHA (only when cwd has a .git directory).

If none resolve, gen_ai.agent.version is omitted. The platform Versions tab then stays empty (it filters agent_version != '').

LiveKit job processes: with python agent.py dev (or spawned job_proc workers), __main__ is LiveKit’s IPC entrypoint, not your agent file. A module-level __version__ on agent.py often does not resolve. The parent dev watcher also skips configure() via _is_livekit_dev_watch_parent, so instrumentation runs in the child where __main__ is not your entrypoint. Example dirs like examples/livekit/restaurant-agent usually have no .git, so the git SHA fallback also fails.

Reliable patterns for LiveKit:

configure(agent_id="restaurant-agent", version="0.1.0")

or set PARLOT_AGENT_VERSION=0.1.0 at deploy/runtime.

2. await ctx.connect() before session.start() (required)

Call explicitly in your entrypoint. Parlot patches JobContext.connect to:

• Refresh platform.ref.room_sid and room metadata on the live session span (paste-search and platform UI linking).
• Start Room Composite Egress to Cloudflare R2 when recording is enabled.

async def entrypoint(ctx: JobContext):
    await ctx.connect()

    session = AgentSession(...)
    await session.start(agent=..., room=ctx.room)

Without ctx.connect() you may still see basic telemetry — turns, handoffs, session close — if you pass room=ctx.room to session.start(). You will not get:

• Room audio recording (egress to R2)
• A reliable room_sid on the session span for paste-search and UI linking

Always call ctx.connect() in production agents and in all Parlot examples.

Environment

Set PARLOT_ENDPOINT and PARLOT_API_KEY. The API key is org-scoped — mint it in Parlot Settings → API Keys (shown once on create). Recording needs Settings → Recording (or configure(record=…) / job metadata) plus agent LIVEKIT_* credentials. Recordings upload to R2 and confirm via lazy R2 reconcile when the session is opened.

Recording policy (precedence: job metadata → configure(record=…) → Settings → Recording via bootstrap):

• UI: Parlot → Settings → Recording (per-agent toggles + globs for unseen agents)
• Code: configure(record=True), configure(record=False), or configure(record=["my-agent*"])
• Dispatch: job metadata { "record": true|false }

Session application logs (Python logging, default on) use the same control path via Settings → Logs / configure(capture_logs=…) / job capture_logs. Not print(). Treat content like stdout for PII.

Generative AI content capture (message bodies / tool payloads; default on) is documented in Concepts. Override with configure(capture_genai_content=True|False).

Optional: PARLOT_DIAGNOSTICS=off to disable SDK self-diagnostics (export/handler failures reported to Parlot). Default is on; metadata only, no conversation content. Under sustained outage the buffer drops and circuit-breaks — it does not retry-storm the collector.

Examples

• examples/livekit/livekit-voice — minimal hello-world
• examples/livekit/multi-agent — storytelling handoffs (graph demo)
• examples/livekit/restaurant-agent — greeter → specialist routing
• examples/livekit/hotel-receptionist — boutique-hotel receptionist + seed sample sessions

Development

uv sync --group dev
uv run pytest -q