LLM observability

See every token. Debug any trace.

OpenTelemetry-native tracing for LLM apps. Instrument once and inspect every LLM call, tool call, and RAG retrieval across dev, staging, and production - with cost and latency on every span.

Read the docs $ pip install langwatch

trace · handle_message1.84s · 6 spans

handle_message

1.84s

• RAG Document Retrieval 0.42s
• embed query 0.12s
• vector_search 0.20s
• generate_answer 1.24s
• format_response 0.06s

The new trace explorer.

Ask for the traces you need, visualize them however debugging demands, and save the views you return to, all on top of the observability you already rely on.

Ask AIshow me the top 10 traces with errors

Origin

• Application 1.1K
• Evaluation 1.0K
• Simulation 445
• Gateway 0

Status

• OK 2.2K
• Error 385
• Warning 0

Error message

• Could not connect to Weaviate 95
• 403 Cloud Firestore API not enabled 37
• requires string as left operand 18
• File not found: file_011 stale 11

Waterfall

Flame

Topology

• Sequence 6 spans · 1.84s

Gantt-style timeline. Every span is a row; the bar shows when it started and how long it ran. The fastest way to spot slow steps.

• trace handle_message 1.84s
• rag RAG Document Retrieval 0.42s
• llm embed query 0.12s
• tool vector_search 0.20s
• llm generate_answer 1.24s
• fn format_response 60ms

AI-powered Ask

Ask the explorer in plain English, "show me the top 10 traces with errors", instead of building filters by hand. A faster way to debug.

Richer visualizations

Waterfall, flame charts, span lists, topology graphs, and sequence diagrams for agentic systems, so unexpected tool-call patterns stand out.

Read it your way

Inspect input and output in a pretty view, or as plain text, JSON, or markdown. Your preferred way of looking at the data.

Waterfall + full audit trail

See input/output sequencing and the complete agent-to-agent communication trail, visibility that basic logs simply cannot give you.

Filterable, linked to evals

Filter by model, project, labels, metadata, and origin (application, eval, simulation), then build datasets straight from selected traces.

Your own lenses

Save a view as a lens. Come back to traces and your layout, filters, and columns are exactly how you left them.

And all the observability you already count on: errors in one place, cost, tokens, duration, models, events, evals, and annotations, ready to build graphs from straight off your traces.

Already works with your whole stack.

LangWatch traces are standard OpenTelemetry spans, enriched with semantic attributes (langwatch.span.type, langwatch.inputs, langwatch.outputs). Keep them in LangWatch, or fan them out to Grafana, Datadog, Jaeger, or any OTel backend you already run. Same spans, every backend.

• your OTel backend

Auto-instrumented

Pydantic AI…and anything that speaks OpenTelemetry

Start tracing.

Get a demo Read the docs

Platform

• Agentic AI Testing Run realistic user scenarios against your agent to catch issues before production
• LLM Evaluation Measure response quality and accuracy so you ship agents that hold up in production
• LLM Observability Trace every agent step and monitor cost and latency with full production visibility
• AI Governance New Govern every model, key, and tool. Virtual keys with budgets, routing policies, and a full audit trail
• Prompt Management Version, deploy, and A/B test prompts as code, with full history and GitHub sync
• Voice AI Test and simulate your voice AI agents at scale, before they talk to customers
• LLM Red-teaming Simulated attacks that uncover safety and security gaps in AI agents