Short answer
The best open-source LLM observability tools in 2026 are OpenLIT (Apache 2.0), Langfuse (MIT core), Arize Phoenix (Elastic License 2.0), Comet Opik (Apache 2.0), MLflow (Apache 2.0), Helicone (Apache 2.0), and OpenLLMetry (Apache 2.0 instrumentation). OpenLIT is the most complete Apache-2.0, OpenTelemetry-native option, covering tracing, evals, guardrails, prompts, and GPU monitoring in one self-hosted platform.
Open-source LLM observability lets you keep prompts, completions, and traces on your own infrastructure, avoid per-trace pricing, and change vendors later. The license matters: Apache 2.0 and MIT are OSI-approved, while Elastic License 2.0 is source-available with restrictions.
This guide covers the open-source and source-available tools teams self-host today. We build OpenLIT and list it first; each tool has the same criteria and honest limitations.
1. OpenLIT
Best for: Open-source, OpenTelemetry-native agent and LLM observability in one self-hosted platform
OpenLIT is an open-source AI agent and LLM observability platform built on OpenTelemetry. It combines tracing, cost and token tracking, evals (LLM-as-a-judge), guardrails, Prompt Hub, a secrets Vault, model comparison (OpenGround), coding agent monitoring, and GPU monitoring in one Apache-2.0 platform you self-host on ClickHouse.
Strengths
- Apache 2.0 with no per-trace fees; data stays in your own ClickHouse
- OpenTelemetry-native SDKs for Python, JavaScript, and Go that auto-instrument LLM providers, agent frameworks, vector DBs, and MCP
- Coding agent observability for Claude Code, Codex, and Cursor through the openlit CLI, no SDK needed
- GPU monitoring for NVIDIA, AMD, and Intel next to inference traces
- Exports OTLP to Grafana, Datadog, or any OpenTelemetry backend, so there is no lock-in
Limitations
- Managed cloud is not generally available yet; today you self-host
- Smaller community than Langfuse or LangSmith
- RBAC, audit logs, alerts, and the eBPF zero-code controller are in the paid Enterprise Edition
License: Apache 2.0 · Self-host: Yes (Docker, Kubernetes Helm). Free. · Website
2. Langfuse
Best for: Popular open-source tracing, prompt management, and evals with a managed cloud
Langfuse is one of the most widely adopted open-source LLM engineering platforms, covering tracing, prompt management, datasets, and evaluations, with both self-hosted and cloud options.
Strengths
- Large community and many framework integrations
- Mature prompt management, datasets, and annotation workflows
- Managed cloud available
Limitations
- Some enterprise features are under a commercial license
- No GPU or infrastructure monitoring
License: MIT core + commercial enterprise features · Self-host: Yes · Website · OpenLIT vs Langfuse
3. Arize Phoenix
Best for: Notebook-friendly tracing and evals, with an upgrade path to Arize AX
Arize Phoenix is an AI observability and evaluation tool from Arize, built on OpenTelemetry and OpenInference. Arize AX is the commercial enterprise platform.
Strengths
- Strong evaluation and experiment tooling
- Good fit for data science and notebook workflows
- Clear path to the enterprise Arize AX platform
Limitations
- Elastic License 2.0 is not OSI-approved open source, which matters to some legal teams
- No built-in GPU monitoring or secrets management
License: Elastic License 2.0 (source-available) · Self-host: Yes · Website · OpenLIT vs Arize Phoenix
4. Comet Opik
Best for: Apache-2.0 tracing and evals from the Comet ML team
Opik is an open-source LLM evaluation and tracing platform from Comet, with self-hosted and cloud deployments.
Strengths
- Apache 2.0 license
- Evaluation-focused feature set with experiments and datasets
- Fits teams already using Comet for ML experiment tracking
Limitations
- Limited infrastructure and GPU monitoring
- Less focus on OTLP export to other backends
License: Apache 2.0 · Self-host: Yes · Website · OpenLIT vs Comet Opik
5. MLflow
Best for: Teams already using MLflow or Databricks
MLflow is the open-source ML lifecycle platform. MLflow Tracing and its GenAI evaluation features add LLM and agent observability to the same tool used for experiments and model registry.
Strengths
- Apache 2.0 and very widely deployed
- One tool for classic ML and GenAI experiments
- Managed option on Databricks
Limitations
- LLM observability is one part of a broader ML platform
- No GPU monitoring, guardrails, or secrets management for LLM apps
License: Apache 2.0 · Self-host: Yes (managed on Databricks) · Website
6. Helicone
Best for: Proxy-based request logging and cost tracking with minimal code
Helicone is an open-source LLM observability tool and AI gateway that logs requests by sitting as a proxy between your app and the model provider.
Strengths
- One-line setup by changing the provider base URL
- Request logging, caching, and cost tracking at the gateway
Limitations
- Proxy sits in the request path
- Limited visibility into multi-step agent internals, tool calls, and vector DBs
License: Apache 2.0 · Self-host: Yes · Website · OpenLIT vs Helicone
7. OpenLLMetry (Traceloop)
Best for: OpenTelemetry instrumentation you send to an existing backend
OpenLLMetry is an Apache-2.0 set of OpenTelemetry instrumentations for LLM providers and frameworks, maintained by Traceloop. It is an instrumentation library, not a full platform.
Strengths
- OpenTelemetry-native instrumentation
- Works with any OTLP backend
Limitations
- No built-in UI, evals, prompt management, or storage; you need a backend
- Traceloop platform features are commercial
License: Apache 2.0 (SDK) · Self-host: SDK only; Traceloop platform is commercial · Website · OpenLIT vs OpenLLMetry (Traceloop)
Evaluation criteria
- Agent trace depth
- Does it capture the full trajectory (LLM calls, tool calls, MCP requests, retrieval, and sub-agents) or only individual model requests?
- OpenTelemetry support
- Native OTLP output keeps your instrumentation portable. Proprietary SDKs make switching tools expensive later.
- Evaluations
- LLM-as-a-judge, programmatic checks, and human review on production traces, plus offline evals you can run in CI.
- Cost and token tracking
- Per request, per user, per model, and per agent run, with custom model pricing.
- License and self-hosting
- OSI-approved open source (Apache 2.0, MIT) versus source-available or closed. Matters for data residency and regulated teams.
- Prompt management and guardrails
- Versioned prompts you can roll back without redeploying, and runtime checks for prompt injection, PII, and unsafe output.
Methodology: we compared public documentation, licenses, and repositories for each tool as of November 18, 2026. OpenLIT is our product. Pricing changes often, so we do not list prices; check each vendor's site. Spot an error? Email [email protected].