Integrations
Choose the smallest Noveum Trace integration for your application stack.
Instrument the complete application boundary and the model, retrieval, tool, and agent operations inside it. Use a framework adapter when one matches the application. Use the Python context managers for custom code or unsupported frameworks.
Choose an integration
| Application | Integration | Start here |
|---|---|---|
| Custom Python or provider client | Context managers for explicit trace and span boundaries | Custom Python |
| LangChain | Runtime callback handler on the top-level runnable | LangChain |
| LangGraph | Runtime callback handler on the compiled graph | LangGraph |
| CrewAI | Global event listener | CrewAI |
| LiveKit Agents | Session tracing with optional STT and TTS wrappers | LiveKit |
| Pipecat | Pipeline observer and task handlers | Pipecat |
Every guide assumes the application and its framework already run correctly. Add only the tracing package and adapter code; do not replace provider, framework, or transport configuration that already works.
Shared requirements
Configure these values before starting instrumented work:
| Variable | Purpose |
|---|---|
NOVEUM_API_KEY | Authenticates trace delivery for the organization. |
NOVEUM_PROJECT | Groups traces for one application or agent. |
NOVEUM_ENVIRONMENT | Separates production, staging, development, and other traffic. |
NOVEUM_SERVICE_VERSION | Identifies the prompt, model, tool, retrieval, and routing behavior that produced the trace. |
Review payload capture before enabling tracing in production. Prompts, model output, retrieval context, tool arguments, transcripts, and audio can contain personal, regulated, or confidential data. Never send credentials, authorization headers, or private keys.
Verify before evaluation
Run representative successful, failed, tool-using, and retrieval-heavy requests. Confirm that each trace covers the complete execution and contains the evidence required by the capture contract.
For the underlying API, use the Python SDK reference. For applications that cannot run Python, use REST trace ingestion.