Bento Documentation

If your app uses Google ADK, Bento captures every model call, tool call, and agent run for you. No per-call code.

import bentolabs_sdk as bento

bento.init(
    user_id=lambda: get_current_user_id(),
    session_id=lambda: get_current_session_id(),
)
bento.instrument()

That’s the whole integration. Run your app and the dashboard fills in immediately.

Not on ADK? For raw LLM SDKs (OpenAI, Anthropic, Bedrock, Vertex), use bento.track_ai. For a framework that already emits OpenTelemetry (LangChain, LlamaIndex, Pydantic AI, or a TypeScript framework like the Vercel AI SDK or Mastra), see Export from any framework. All of these mix freely in the same app.

Install

pip install "bentolabs-sdk[adk]"

The [adk] extra pulls in the OpenInference tracer for ADK. The base wheel stays small.

`bento.instrument()`

Call this once at app startup, after bento.init(...) and before any LLM call.

bento.instrument()           # auto-detects ADK
bento.instrument("adk")      # explicit

With no argument, Bento activates ADK if it’s installed and returns "adk", or returns None when the [adk] extra is missing. Passing "adk" activates ADK by name. An unknown name raises ValueError; a missing [adk] extra logs a warning and returns None. The call is safe to repeat: a second call with the same target is a no-op, and a second call with a different target logs a warning and is ignored. bento.shutdown() removes every active integration on the way out.

Identity at init

To stamp every captured span with user_id, session_id, and tags without threading them through your code, pass getters to bento.init(...):

bento.init(
    user_id=lambda: get_current_user_id(),
    session_id=lambda: get_current_session_id(),
    tags=lambda: ["env:prod"],
)

A getter is any zero-argument callable. Bento calls it on every span and writes the result. Three common patterns:

# FastAPI / Starlette
bento.init(user_id=lambda: Request.scope["state"].user_id)

# Your own ContextVar
from contextvars import ContextVar
current_user: ContextVar[str | None] = ContextVar("current_user", default=None)
bento.init(user_id=current_user.get)

# Flask / Django
bento.init(user_id=lambda: g.user.id if g.user.is_authenticated else None)

If a getter raises, Bento drops the field for that span and your app keeps running. Telemetry never breaks the host. For more, see Configuration → Identity getters.

Add custom events

For ad-hoc events alongside the integration, wrap each one with bento.track_ai: a custom event before or after a model call, or an event for a step ADK doesn’t emit. The integration’s spans and your manual ones land in the same dashboard. If you need a single trace to span multiple LLM and tool calls within one turn, see Trajectories (advanced).

Reference

Function	Effect
`bento.instrument()`	Auto-detect and activate ADK. Returns `"adk"` or `None`.
`bento.instrument("adk")`	Activate by name. Raises `ValueError` on unknown names.
`bento.uninstrument()`	Remove every active integration. Returns the names removed.
`bento.uninstrument("adk")`	Remove ADK specifically.

Manual tracking

bento.track_ai for SDKs without an integration.

Configuration

init(), identity getters, env vars, lifecycle.

Trajectories

Group multi-step work into one trace.

OTel transport

Wire Bento into an existing OpenTelemetry pipeline.

Get started

Integrations

Manual tracking

Advanced

Google ADK

Install

`bento.instrument()`

Identity at init

Add custom events

Reference

See also

Manual tracking

Configuration

Trajectories

OTel transport

​Install

​bento.instrument()

​Identity at init

​Add custom events

​Reference

​See also

Manual tracking

Configuration

Trajectories

OTel transport

Install

`bento.instrument()`

Identity at init

Add custom events

Reference

See also