Overview
Langfuse (31k+ stars) is LLM-native observability with trace + prompt management + evaluation as a stack. SigNoz (31k+ stars) is an OpenTelemetry-native general APM that treats LLM workloads as one of many. We compare positioning, protocol support, data model, extensibility, and typical scenarios.
TypeScript · NOASSERTION
Open-source LLM engineering platform providing tracing, evaluations, prompt management, and dataset management with integrations for LangChain, OpenAI, Anthropic, and more.
| Best for | Langfuse |
|---|---|
| Product positioning | LLM-specific observability: trace, span, generation, prompt, and score are all first-class. UI is purpose-built for LLM debugging (slice by prompt, view feedback per user). |
| Protocol support | OpenTelemetry plus dedicated Python / JS SDK plus LangChain / LlamaIndex integrations. LLM-as-judge and heuristic eval are built in. |
| Data model | LLM-centric: traces contain generations (model calls), tools (tool calls), and scores (evaluations). Prompt versioning and user feedback are first-class. |
| Extensibility | Prompt versioning, LLM-as-judge, custom scores, webhooks. Enough but not flexible. |
| Best fit | Pure LLM products (ChatGPT-style app, agents, RAG) that need trace + prompt debugging + eval loop. First choice for LLM teams. |
| Metric | Langfuse |
|---|---|
| Stars | 33.2k |
| Forks | 3.6k |
| Language | TypeScript |
| License | NOASSERTION |
| Last commit | August 16, 2026 |
Choose based on your primary workflow, language ecosystem, and integration needs. Review each project's documentation and recent GitHub activity before adopting it in production.
Yes. A common pattern is SigNoz as the unified observability stack (trace plus metrics plus logs) and Langfuse as the LLM-specific UI (prompt debugging plus eval). Both support OpenTelemetry export, so a single collector can route data to different backends.
Pick Langfuse for pure LLM products: trace plus prompt management plus eval plus user feedback in one place. Pick SigNoz for mixed workloads (traditional services included): unified observability costs less, and you can still trace LLM data but you organize attributes yourself.
ClickHouse handles high-cardinality queries extremely fast, which fits LLM use cases that slice by user_id, prompt_id, or model. Langfuse uses ClickHouse for OLAP traces plus Postgres for OLTP; SigNoz is ClickHouse end-to-end and lighter to operate.
No — anything OpenTelemetry-instrumented can send data. But the UI is LLM-centric. If you only run LLM applications, it is enough. If you also need to monitor API latency plus DB queries plus LLM in one stack, SigNoz is more general.