Skip to main content
The Bento Python SDK sends analytics about your AI agent runs to Bento, where you can see what your app does in production and improve what fails.

Set up with an AI coding tool

Install Bento’s Agent Skills so your AI coding tool knows how to wire up the SDK. Copy this prompt into Claude Code, Cursor, Codex, OpenCode, or Windsurf:
The skill carries the full playbook: discovering your framework, installing the SDK, wiring up bento.instrument() or bento.track_ai, and verifying the first trace lands. Prefer to do it by hand? The rest of this page walks the same steps manually.

Requirements

Python 3.10 or higher.

Install

If you use Google ADK, install the extra so Bento can capture every call automatically:
The base wheel ships its own OpenTelemetry dependencies. You don’t need to install opentelemetry-api or opentelemetry-sdk separately.

Set your API key

Sign in at platform.bentolabs.ai and create a workspace. Copy the API key from settings. Keys start with bl_pk_ and the prefix is validated by the SDK, so a clearly-bad key fails before any network request goes out.
In production, set the key via environment variable. For local development or notebooks, pass it in code: bento.init(api_key="bl_pk_...").

Send a test trace

Pick the one that matches your stack. Google ADK:
Anything else:
Open platform.bentolabs.ai. The trace lands within seconds.

Next

Export from any framework

Already on LangChain, LlamaIndex, or a TS framework? Point its exporter at Bento.

Integrations

Trace Google ADK with one line.

Manual tracking

Wrap OpenAI / Anthropic / Bedrock / Vertex calls.

Configuration

Identity getters, env vars, lifecycle.

Custom properties

Tag spans with experiment IDs, feature flags, anything.