Langtrace

Langtrace is an open-source observability and evaluation platform purpose-built for LLM applications, AI agents, and retrieval-augmented generation (RAG) pipelines. It provides detailed distributed tracing, cost analytics, and quality evaluation capabilities that help engineering teams understand exactly what their AI systems are doing in production, how much they cost, and how well they perform.

At its core, Langtrace is built natively on the OpenTelemetry standard, which means every trace and span it generates conforms to OTLP conventions and can be exported to any compatible backend — Grafana, Datadog, Signoz, or your own collector. This vendor-neutral approach sets it apart from observability tools that lock telemetry into proprietary formats. For platform teams already running OpenTelemetry infrastructure for microservices, Langtrace slots into the existing stack rather than creating a parallel silo.

The platform auto-instruments 8 major LLM providers (OpenAI, Anthropic, Google Gemini, Cohere, Groq, Mistral, Perplexity, and Ollama) and over 10 orchestration frameworks and vector databases including LangChain, LlamaIndex, LangGraph, CrewAI, DSPy, AutoGen, Pinecone, Chroma, Weaviate, and Qdrant. Instrumentation requires just two lines of code — import the SDK and call init — after which every LLM call, tool invocation, embedding query, and vector retrieval is captured automatically with full prompt and completion content, token counts, latency, and cost.

Cost tracking is a first-class feature. Dashboards aggregate spend by model, user, project, prompt template, and time window, making it straightforward to identify which features or tenants are driving the largest portion of an AI bill. Teams report using this data to set budget alerts, negotiate model pricing, and justify optimization investments to finance stakeholders.

For evaluation and quality management, Langtrace lets teams promote production traces into curated datasets, annotate them with human feedback, run prompt experiments across model versions, and score outputs using built-in evaluators for accuracy, faithfulness, toxicity, and custom metrics. This closes the loop between observability and iteration — instead of treating monitoring and evaluation as separate workflows, teams can move from a suspicious trace to a scored experiment in a few clicks.

The self-hosted deployment option is a significant differentiator for regulated industries. The server is AGPL-3.0 licensed while the SDKs are Apache-2.0, and a Docker Compose file launches the full stack (server, Postgres, ClickHouse) in minutes. Healthcare, finance, and government teams that cannot send raw prompts to third-party SaaS providers can run Langtrace entirely within their own VPC while maintaining standard OpenTelemetry compatibility.

The managed Cloud offering starts with a free tier (50,000 traces/month, 30-day retention), scales through a Pro plan at $59/month (up to 1 million spans included, ~$0.20 per 1,000 additional spans, 90-day retention), and offers custom Enterprise agreements with single-tenant deployment, SOC2 documentation, and SLAs. This tiered approach makes Langtrace accessible to individual developers prototyping an agent as well as platform teams running production workloads at scale.