tuner-livekit-sdk

Automatically ingest LiveKit Agents session data into the Tuner observability API.

Drop it into your entrypoint and every session — transcript, tool calls, timing metrics, usage, cost — is submitted to Tuner when the call ends. No manual event wiring required.

Features

• Zero-friction wiring — one line after AgentSession creation; the plugin hooks itself into session lifecycle events.
• Rich per-turn timing — STT/LLM/TTS latency, end-of-turn delay, and end-to-end latency captured per transcript segment.
• Tool call tracking — function calls and their results are merged into the timeline with real timing data.
• LangGraph / LangChain support — via the bundled tuner-langchain package, capture node transitions and tool calls straight from your graph.
• SIP simulation correlation — match Tuner-initiated simulation calls to your production SIP trunk.
• Cost calculation — plug in your own pricing function; Tuner reports cost per call.
• Resilient delivery — configurable timeouts and retries on transient failures.

Table of Contents

• Installation
• Quickstart
• Configuration
• Options
• Data captured
• Trace view (OpenTelemetry)
• LangGraph / LangChain observability
• Simulation correlation (SIP)
• Privacy & data handling
• Requirements
• Support
• Development
• License

Installation of the Library into your Livekit project

pip install tuner-livekit-sdk

Quickstart

Set credentials via environment variables:

export TUNER_API_KEY="tr_api_..."
export TUNER_WORKSPACE_ID="123"
export TUNER_AGENT_ID="my-agent"

Then drop the plugin in right after creating your AgentSession:

from tuner import TunerPlugin

async def entrypoint(ctx: JobContext):
    session = AgentSession(...)
    TunerPlugin(session, ctx)   # wires itself automatically
    await session.start(...)

That's it. The plugin listens to session events and submits call data to Tuner when the session ends.

Configuration

Environment variables

Variable	Required	Description
TUNER_API_KEY	✅	Bearer token (starts with tr_api_)
TUNER_WORKSPACE_ID	✅	Integer workspace ID
TUNER_AGENT_ID	✅	Agent identifier from Tuner Agent Settings
TUNER_BASE_URL	—	API base URL (default: https://api.usetuner.ai)

Credentials from code

Pass credentials directly instead of (or to override) environment variables:

TunerPlugin(
    session, ctx,
    api_key="tr_api_...",
    workspace_id=123,
    agent_id="my-agent",
)

Options

Call type

By default the plugin auto-detects the call type (phone_call for SIP participants, web_call otherwise). Override it explicitly:

TunerPlugin(session, ctx, call_type="phone_call")
TunerPlugin(session, ctx, call_type="web_call")

Recipient (callee)

Pass the phone number or SIP URL of the called party when your agent initiates or routes outbound calls. This field is not auto-collected — supply it explicitly when known:

# E.164 phone number
TunerPlugin(session, ctx, recipient="+15551234567")

# SIP URI
TunerPlugin(session, ctx, recipient="sip:alice@example.com")

recipient is optional. If omitted it is simply not included in the call record.

Recording URL

Tuner requires a recording_url for every call. If you don't provide a resolver the plugin logs a warning and submits "pending" as a placeholder:

# Static URL
async def my_resolver(room_name: str, job_id: str) -> str:
    return f"https://cdn.example.com/recordings/{job_id}.ogg"

TunerPlugin(session, ctx, recording_url_resolver=my_resolver)

# LiveKit Egress → S3
async def egress_resolver(room_name: str, job_id: str) -> str:
    url = await my_egress_db.get_recording_url(room_name)
    return url or "pending"

TunerPlugin(session, ctx, recording_url_resolver=egress_resolver)

Cost calculation

Provide a callable that receives a UsageSummary and returns the call cost in cents:

def calculate_cost(usage) -> float:
    llm_cost  = usage.llm_prompt_tokens     * 0.000_003
    llm_cost += usage.llm_completion_tokens * 0.000_015
    tts_cost  = usage.tts_characters_count  * 0.000_030
    stt_cost  = usage.stt_audio_duration    * 0.000_006
    total_dollars = llm_cost + tts_cost + stt_cost
    return round(total_dollars * 100, 2)  # dollars -> cents

TunerPlugin(session, ctx, cost_calculator=calculate_cost)

Extra metadata

Attach arbitrary key-value data to every call record:

TunerPlugin(
    session, ctx,
    extra_metadata={
        "env": "production",
        "region": "us-east-1",
        "deployment": "v2.3.1",
    },
)

Retry and timeout

TunerPlugin(
    session, ctx,
    timeout_seconds=15.0,   # per-request timeout (default: 30.0)
    max_retries=5,          # retries on 5xx / 429 / network errors (default: 3)
)

Agent version tracking

Track which version of your agent handled each call — useful when you update a prompt, swap a model, or change your pipeline:

AGENT_VERSION=42 python agent.py start

Tuner reads it automatically. Bump the number on every deployment.

Override in code (takes priority over the env var):

TunerPlugin(session, ctx, agent_version=42, ...)

Disable the plugin

Useful for local development or test environments:

import os

TunerPlugin(
    session, ctx,
    enabled=os.getenv("ENV") == "production",
)

Data captured

Each item in session.history is mapped to a transcript segment (role, text, start_ms, end_ms, metadata). For user / agent segments, metadata includes the following per-turn timing and quality fields:

Field	Type	Description
id	string	LiveKit ChatContext item ID
interrupted	bool	Whether this turn was interrupted before completing
transcript_confidence	float | null	STT confidence score for user turns
stt_node_ttfb	int (ms) | null	Time to obtain the transcript after the user stopped speaking
eou_delay	int (ms) | null	Time between end of user speech and the decision to end their turn
llm_node_ttft	int (ms) | null	Time for the LLM node to return its first token
tts_node_ttfb	int (ms) | null	Time for the TTS node to return its first audio chunk
e2e_latency	int (ms) | null	Time from end of user speech to the agent beginning its response

All timing fields above are taken directly from LiveKit's own per-turn metrics (EOUMetrics, STT/LLM/TTS metrics) and passed through unchanged — the SDK does not recompute them.

Tool calls (agent_function / agent_result roles) and, when LangGraph/LangChain instrumentation is enabled, node_transition segments are interleaved into the same timeline, sorted by start_ms.

Trace view (OpenTelemetry)

LiveKit Agents already emits OpenTelemetry spans for every session — STT, LLM, TTS, tool calls, end-of-utterance decisions. The plugin forwards them to Tuner, which shows them as a trace tree on the call details page: what ran, in what order, and how long each step took.

Setup

Install the extra:

pip install 'tuner-livekit-sdk[traces]'

That is the whole setup. You do not need to enable telemetry in LiveKit, build a TracerProvider, choose an exporter, or set any OTEL_* environment variable — the plugin does all of that using the API key and base URL you already pass it, and tags every span with the call id so the trace lands on the right call.

# No tracing-specific arguments needed.
TunerPlugin(session, ctx, api_key=..., agent_id=...)

Turning it off

TunerPlugin(session, ctx, forward_traces=False)

Tracing is also a no-op — with a debug log, never an error — when the traces extra is not installed. It can never fail a call.

If you already export traces somewhere else

Your setup is preserved. When the plugin finds a TracerProvider you configured, it adds Tuner as an additional destination rather than replacing it, so your existing backend keeps receiving the same spans.

What is stored

The span tree — names, timings, parent relationships, status — plus metrics and configuration: model names, time-to-first-byte, token counts, turn latency, endpointing settings, and tool names and outcomes.

No conversation content. Transcripts, LLM inputs and outputs, TTS text, system prompts, and tool call arguments and results are all dropped on arrival and never stored. The filter is an allowlist, so anything not explicitly approved is discarded by default.

Roughly 30 KB per call, for a five-minute call with twenty turns.

LangGraph / LangChain observability

tuner-langchain ships as a dependency of tuner-livekit-sdk, so no separate install step is needed. If your agent uses LangGraph or LangChain as the orchestration layer, wire it in with wrap_graph() or wrap_chain(). Each returns a drop-in replacement for the graph/chain you pass in — hand it straight to LLMAdapter, no callbacks to wire up:

from tuner import TunerPlugin
from livekit.plugins import langchain

plugin = TunerPlugin(session, ctx)

llm = langchain.LLMAdapter(
    plugin.wrap_graph(my_graph),
    stream_mode="messages",
)

Use wrap_chain() instead of wrap_graph() for a plain (non-graph) LangChain runnable.

To limit what data is forwarded to Tuner, pass a CaptureConfig:

from tuner import TunerPlugin
from tuner_langchain import CaptureConfig

plugin = TunerPlugin(session, ctx)
wrapped_graph = plugin.wrap_graph(
    my_graph,
    capture=CaptureConfig(
        tool_inputs=False,
        node_instructions=False,
    ),
)

Simulation correlation (SIP)

Tuner simulations dial into your agent through the same SIP trunk that handles production phone calls. Tuner sends an X-Correlation-Id header on the SIP INVITE. The SDK reads it from the SIP participant and forwards it as sip_headers, so Tuner can link the call to its simulation run. The SDK also forwards LiveKit's sip.callIDFull attribute as sip_call_id.

Both are captured automatically. You don't need to read participant attributes or pass sip_call_id yourself.

This section covers the SDK side only. For LiveKit platform setup (SIP URI, inbound trunk, dispatch rule, Tuner SIP settings), see:

→ docs.usetuner.ai/docs/api-and-integrations/connecting-to-livekit/simulation-setup

Requirements

• tuner-livekit-sdk newer than 0.1.9 for X-Correlation-Id forwarding (sip_call_id capture needs >= 0.1.5).
• Your inbound trunk must map SIP X-headers to participant attributes: set include_headers to SIP_X_HEADERS (or SIP_ALL_HEADERS). LiveKit then exposes the header as the sip.h.x-correlation-id attribute.

⚠️ Trunk setting: the default, SIP_NO_HEADERS, maps no headers. The call still reports, but without the correlation header, and nothing warns you. See LiveKit's SIP participant attributes docs.

How capture works

The plugin reads the SIP participant's attributes when the participant joins, again whenever LiveKit updates them (participant_attributes_changed), and once more at shutdown. The first non-empty value wins, so a later SIP participant (for example a transfer leg) can't replace the caller's identifiers.

Overriding sip_call_id

Pass sip_call_id= only to override the captured value:

TunerPlugin(session, ctx, sip_call_id="my-own-call-id")

Privacy & data handling

The plugin submits the following to the Tuner API for each call: the plain-text transcript, per-turn timing metadata (see Data captured), tool call names/arguments/results, token/character/audio usage counts, and any extra_metadata you configure. It does not access raw audio, video, or room recordings directly — recordings are only referenced by the URL your recording_url_resolver returns.

Use CaptureConfig (see LangGraph / LangChain observability) to exclude tool inputs or node instructions from LangGraph/LangChain traces, and enabled=False to disable submission entirely for local development or test environments.

Requirements

• Python ≥ 3.10
• livekit-agents >= 1.4
• tuner-livekit-sdk >= 0.1.5 for sip_call_id; newer than 0.1.9 for X-Correlation-Id forwarding (see Simulation correlation (SIP))
• aiohttp >= 3.9
• tuner-langchain >= 0.1.1 (installed automatically as a dependency; used by wrap_graph() / wrap_chain())

Support

• Docs: docs.usetuner.ai
• Issues / feature requests: GitHub Issues
• Email: support@usetuner.ai
• Release notes: CHANGELOG.md

Development

Set up a local environment:

uv sync --dev
source .venv/bin/activate

Run the test suite:

uv run pytest -v

Publishing to PyPI

pip install build twine
python -m build
twine upload dist/*

License

MIT