Install

pip install auditant

import auditant
auditant.init(api_key="ak_…", log_id="tenant_acme/prod", agent_id="underwriter")

With OpenTelemetry present, init() registers a span processor and activates whichever supported OpenInference instrumentors are installed — LangChain (which is what sees LangGraph), CrewAI, LlamaIndex, OpenAI, Anthropic and Google GenAI — with no changes to your agent's logic. When the OpenAI Agents SDK is importable, its trace-processor bus is registered on too. Install the extras for the frameworks you run:

pip install "auditant[openinference]"     # every supported auto-instrumentor
pip install "auditant[langchain]"         # just LangChain / LangGraph
pip install "auditant[crewai]"            # just CrewAI
pip install "auditant[llama-index]"       # just LlamaIndex
pip install "auditant[google-genai]"      # Gemini through Google GenAI
pip install "auditant[openai-agents]"     # the OpenAI Agents SDK — init() registers its trace processor
pip install "auditant[gateway]"           # the LiteLLM gateway callback

Without OpenTelemetry, init() still gives you a session on the chain (lifecycle + heartbeats, so a quiet agent is distinguishable from a dead emitter), the manual record() API, and the synchronous decide() policy check. The handle's activated list states exactly what turned on — and an instrumentor installed for a framework that is not (CrewAI's, say, with no CrewAI) is reported as not activated. An audit SDK that implies more coverage than it has would be the product contradiction.

npm i auditant

import { init } from "auditant";
const auditant = init({ apiKey: "ak_…", logId: "tenant_acme/prod", agentId: "support-bot" });

const result = await generateText({
  model: openai("gpt-4o"),
  experimental_telemetry: { isEnabled: true },   // the AI SDK's own switch
  // …
});
await auditant.flush();   // serverless: before the response returns

The processor attaches to whatever tracer provider your app already runs — this SDK neither installs nor configures OpenTelemetry, because two tracing setups fighting is worse than one. flush() is awaitable and coalesced: a Vercel function freezes the instant the response returns, and an event sitting in a queue at that moment never happened. The AI SDK's own forceFlush routes into it.

Events appear on your dashboard's Overview within seconds of the first run. Model calls carry observed token counts (never invented prices — attributionComplete: false until a pricing table says otherwise), tool calls carry input/output hashes, and everything shares a session so the whole arc reads as one record. Next: verify a bundle yourself, or wire an adapter if your agents run behind a gateway or the OpenAI Agents SDK.