Observability
noveum-trace records instrumented LLM calls, tool invocations, retrievals, and agent steps as connected spans, with token, cost, and latency metadata when available. Use supported framework callbacks and custom operations to capture the execution context your team needs.
Production
trace volume at scale
Hierarchical
agent and workflow traces
Guided
setup through first trace
01: Capabilities
Supported integrations include LangChain, LangGraph, LiveKit, Pipecat, and CrewAI. Each integration records the callbacks or wrapped operations it receives; verify the resulting spans for your framework version and workflow.
noveum-trace stores supplied token and reported-cost metadata with each span. When reported cost is absent, analysis can estimate it from a recognized provider and model using the current cost registry.
Conversation threads, agent registries, and workflow graphs keep recorded multi-agent operations legible - who did what and in which order.
Audio and image artifacts upload alongside spans: the raw material that voice and vision evals need.
The SDK batches and retries exports and is designed to fail softly when telemetry delivery fails. Validate its behavior and overhead in your own runtime.
Traces, spans, attributes, and events that map cleanly onto OTel concepts, with a purpose-built wire format for LLM workloads.
02: Integration
Initialize the SDK, add the supported integration for your stack, and verify the first trace in Explorer. Our engineers can guide the integration during onboarding.
from noveum_trace.integrations.langchain import ( NoveumTraceCallbackHandler,) handler = NoveumTraceCallbackHandler() agent.invoke( {"input": question}, config={"callbacks": [handler]},)What gets captured
Python 3.9–3.13 · OpenTelemetry-aligned · MCP server included
03: Under the hood
A columnar trace store, an open-source Python SDK, an embedded cost registry, and an MCP server - the concrete pieces that make agent observability queryable.
Storage engine
A ClickHouse columnar backend designed for high-volume trace aggregation and storage efficiency.
Open-source Python SDK
Python `noveum-trace` (3.9–3.13) - instrument with context managers, one init call to start streaming spans.
Model & cost registry
An embedded registry covering 40+ models across 11 providers: OpenAI (gpt-4o, o1, o3, gpt-5), Anthropic Claude 3/3.5/4.5, Google Gemini 1.5/2.0/2.5, Cohere, Mistral, Meta Llama, DeepSeek, xAI Grok, Perplexity, Amazon Bedrock, and AI21.
Per-span economics
Captured token counts and reported cost remain attached to the relevant span. Analysis can estimate missing cost from recognized provider and model metadata using the current registry.
Native integrations
Documented integrations for LangChain, LangGraph, LiveKit, Pipecat, and CrewAI, with capture coverage determined by the callbacks or wrapped operations delivered.
MCP server
An MCP server generated from the current OpenAPI specification exposes supported Noveum actions to your own agents, alongside curated resources and guided prompts.
The loop
noveum-trace supplies production evidence for dataset-based evaluation and NovaPilot analysis. NovaSynth can test a configured endpoint before or after deployment, while NovaGuard policies are enforced by a connected runtime.
Trusted in production by enterprise and high-growth teams