Choose an integration

Find the integration that fits how your agent is built, in Python or TypeScript.

Judgment ingests OpenTelemetry spans, and each integration is a different way to create them and hand them to the Judgment exporter. Every page in this section assumes you’ve initialized the SDK as in Set up tracing, with JUDGMENT_API_KEY and JUDGMENT_ORG_ID set.

Pick the row for whatever does the most work in your agent. That’s usually the framework running the tool loop rather than the model client underneath it.

What creates your spans	Python	TypeScript	Page
You call the OpenAI client directly	`wrap()`	`wrap()`	Model clients
You call Anthropic, Together, or Google GenAI directly	`wrap()`	OpenTelemetry instrumentation	Model clients
Bedrock, Vertex AI, Groq, Ollama, LiteLLM, other providers	OpenTelemetry instrumentor	OpenTelemetry instrumentation	Model clients
LangGraph or LangChain	OpenInference instrumentor	Send OTLP	LangGraph and LangChain
OpenAI Agents SDK	OpenInference instrumentor	Send OTLP	OpenAI Agents SDK
Claude Agent SDK	`setup_claude_agent_sdk()`	Send OTLP	Claude Agent SDK
Claude Managed Agents	Not applicable	`wrapAnthropicManagedAgents()`	Claude Agent SDK
Vercel AI SDK	Not applicable	AI SDK telemetry	Vercel AI SDK
Google ADK, Pipecat, LiveKit Agents, Pydantic AI	Install Judgment as the global provider	Send OTLP	Frameworks with built-in OpenTelemetry
FastAPI, Flask, Django, httpx, SQLAlchemy, Redis, Celery, and other libraries	`Tracer.registerOTELInstrumentation()`	`Tracer.registerOTELInstrumentation()`	OpenTelemetry instrumentation libraries
An OpenTelemetry SDK or Collector you already run, any language, OpenClaw	OTLP over HTTP	OTLP over HTTP	Send OTLP directly
Logfire, LangSmith, Braintrust, or Langfuse	Dual export	Dual export	Keep your existing tool

“Send OTLP” in the TypeScript column means Judgment has no TypeScript helper for that framework. If the framework emits OpenTelemetry spans through the global provider, call JudgmentTracerProvider.installAsGlobalTracerProvider() after Tracer.init() so they go through the SDK to Judgment. Otherwise, point its exporter at Judgment.

What’s one line and what’s real setup

wrap(), setup_claude_agent_sdk(), and, for most libraries, Tracer.registerOTELInstrumentation(...) are one line after Tracer.init.
Frameworks that only read the global OpenTelemetry provider need one extra call, JudgmentTracerProvider.install_as_global_tracer_provider() in Python or JudgmentTracerProvider.installAsGlobalTracerProvider() in TypeScript. Tracer.init doesn’t install it for you.
Raw OTLP and dual export are real setup. You build the exporter, set three headers, and need the project ID instead of the project name.

Rules for every integration

Record each model call once. If a framework integration already records the model call, don’t also wrap() the client or add a provider instrumentor. You’d get duplicate model spans, tokens, and cost.
Keep one project name per agent. The name creates the project if it doesn’t exist (Set up tracing explains the lookup). Use environment= for dev, staging, and production.
Flush before short-lived processes exit. Scripts, Lambda handlers, and CLI tools should call Tracer.force_flush() (Python) or await Tracer.shutdown() (TypeScript) before they return.