Framework Adapters

retrace.init() auto-patches the OpenAI / Anthropic / Gemini SDKs, so any framework that calls them under the hood is already captured. Adapters add the framework-specific structured spans (tool / retrieval / chain / agent) the detectors rely on.

One-line init

import retrace
retrace.init()   # reads RETRACE_API_KEY, auto-patches installed providers, auto-starts a trace

import { init } from "retrace-sdk";
init();          // same, for Node

No @record / startSpan boilerplate needed for the common case.

LangChain & LangGraph

import retrace
from retrace.adapters.langchain import RetraceCallbackHandler

retrace.init()
agent.invoke(input, config={"callbacks": [RetraceCallbackHandler()]})   # works for LangGraph too

import { init } from "retrace-sdk";
import { createLangChainHandler } from "retrace-sdk/adapters/langchain";

init();
const cb = await createLangChainHandler();
await app.invoke(input, { callbacks: [cb] });

CrewAI

import retrace
from retrace.adapters.crewai import instrument_crew

retrace.init()
crew = instrument_crew(Crew(agents=[...], tasks=[...]))
crew.kickoff()

Vercel AI SDK (and Mastra)

The AI SDK talks to providers through its own packages, so use the step adapter (it records the LLM step and the tool calls/results):

import { init } from "retrace-sdk";
import { retraceOnStepFinish } from "retrace-sdk/adapters/vercel-ai";

init();
await generateText({ model, prompt, tools, onStepFinish: retraceOnStepFinish() });

OpenAI Agents SDK

No adapter needed — the Agents SDK calls the OpenAI SDK, which init() already patches. Just call retrace.init() before running your agent.

Span shapes align with the detectors

Adapters emit tool_call (args → schema validation), verbatim tool_result (→ tool-output hallucination + loop detection), action/retrieval (→ replay divergence), and reasoning spans — so error detection works on framework traces with no extra wiring.

Framework Adapters