Best / LLM observability
Best open-source LLM observability
The best self-hosted LLM-observability split in the c17.io catalog is Langfuse for MIT-core traces and evals (the published LangSmith stack — skip ee/), Helicone for an Apache-2.0 observability proxy, Phoenix when you accept Elastic License 2.0 (not OSI; you may not offer it as a hosted service to third parties), and MLflow when the job is Apache-2.0 experiment tracking instead of Weights and Biases. Serving stays on /best/model-serving. You still pay tokens to the model vendor.
Open the stackStart on Railway
Ranked picks
1
LangfuseMIT core LLM traces and evals (skip ee/) — the c17 LangSmith path.
When to pick it: Traces and evals on Postgres you control.
36,145 stars in the catalog
2
HeliconeApache-2.0 LLM observability proxy you can self-host — /compare/langfuse-vs-helicone.
When to pick it: A proxy in front of the model API.
6,231 stars in the catalog
3
PhoenixUseful, and not OSI. Elastic License 2.0 forbids offering it as a hosted service to third parties — /compare/langfuse-vs-phoenix.
When to pick it: ELv2 evals if you accept that license.
11,939 stars in the catalog
4
MLflowApache-2.0 experiment tracking — the c17 Weights and Biases path. Training runs, not LLM traces.
When to pick it: Runs, params, and artifacts. /replace/wandb.
28,480 stars in the catalog
How they compare
| Tool | License | Self-host | Best for |
|---|---|---|---|
| Langfuse | MIT (core; skip ee/) | Medium | Traces and evals on Postgres you control. |
| Helicone | Apache License 2.0 | Medium | A proxy in front of the model API. |
| Phoenix | Elastic License 2.0 | Medium | ELv2 evals if you accept that license. |
| MLflow | Apache License 2.0 | Medium | Runs, params, and artifacts. /replace/wandb. |
Related pages
- Langfuse / Helicone stack
- MLflow / Aim stack
- vLLM / LocalAI stack
- Replace LangSmith
- Replace Weights and Biases
- Replace Hugging Face Inference Endpoints
- Replace Replicate
- Langfuse vs Phoenix
- Langfuse vs Helicone
- Helicone vs Phoenix
- MLflow vs Aim
- MLflow vs Langfuse
- All best open-source pages
- All compare pages
- All replace pages
- All stacks
- Tools catalog
FAQ
- Langfuse or Helicone?
- Langfuse if you want traces and evals (the published LangSmith stack). Helicone if you want that Apache-2.0 proxy — /compare/langfuse-vs-helicone.
- Is this Weights and Biases?
- Training runs are MLflow — /replace/wandb and /compare/mlflow-vs-aim. This page is LLM traces first.
- Why is Phoenix last?
- Elastic License 2.0 is not OSI. Do not offer it as a hosted service to third parties — /compare/langfuse-vs-phoenix.
- Are these free?
- Langfuse core is MIT (skip ee/). Helicone and MLflow are Apache-2.0. Phoenix is ELv2. You pay hosting and model tokens. Template users never pay c17.
- How do I deploy it?
- Open /stacks/langfuse-observe. Persist Postgres. Point one app at the ingest URL. Skip ee/.
Stack of the Week
Get one stack a week that replaces a SaaS bill.