redberry-sdk

Records an AI agent's model and tool calls with OpenTelemetry and sends them to the customer's own Redberry install. Redberry uses these records to keep a tamper-evident evidence chain for each agent and to price its risk.

The SDK sends only to the Redberry install in your own cloud account (or to an endpoint you set). It does not send anything to Redberry directly.

Install

pip install redberry-sdk                    # core
pip install "redberry-sdk[providers]"       # plus instrumentation for model providers (OpenAI, Anthropic, ...)
pip install "redberry-sdk[all]"             # providers, agent frameworks and vector databases

Python 3.11 or later.

Quick start

import redberry_sdk

redberry_sdk.init(agent="refund-agent")   # your agent's name, as you will approve it in the Redberry portal

Call init() once, before the agent makes its first model call. Model calls made through supported libraries are then recorded automatically.

Serverless functions (AWS Lambda, Azure Functions, Cloud Run functions) can be frozen as soon as a request returns. Flush before returning:

@redberry_sdk.flush_after
def handler(event, context):
    ...

or call redberry_sdk.flush() yourself.

Temperature. Redberry prices an agent from its sampled behaviour, so the agent's model calls must run at temperature 1.0.

Where records go

On these platforms the SDK finds your install automatically and authenticates with the platform's own identity; no keys are needed:

Elsewhere, set the ingest address explicitly.

Settings can also be placed in a .env file next to the agent. Settings from earlier versions of the SDK, under a different prefix, are still accepted.

Records the SDK could not deliver are counted by redberry_sdk.dropped_spans().

License

Apache License 2.0. See LICENSE and NOTICE.