# Langfuse / Helicone

LLM traces you host. Langfuse core is MIT; Helicone is Apache-2.0 — LangSmith without the hosted observability seat.

**Replaces:** LangSmith

**Tier:** Medium

## Problem

LangSmith is LangChain's hosted trace and eval product. You wanted traces on Postgres you control. Langfuse is MIT outside ee/ (langfuse/langfuse) — skip ee/. Helicone is Apache-2.0 (Helicone/helicone). Phoenix (Arize-ai/phoenix) is Elastic License 2.0 — not OSI; you may not provide it as a hosted service to third parties. This is not LangSmith's every eval SKU. Training runs stay on /replace/wandb. You still pay tokens to the model vendor.

## Stack

- [Langfuse](https://langfuse.com) — MIT core LLM traces and evals (skip ee/)
- [Helicone](https://www.helicone.ai) — Apache-2.0 LLM observability proxy you can self-host

## Estimated SaaS cost saved

| You stop paying | Typical / year | Self-host instead |
|---|---:|---|
| LangSmith (small team) | $1,200 | Langfuse (MIT core) or Helicone |

**About $1,200 / year** on a small team, before you count lock-in.

## Deploy on Railway

Placeholder: Langfuse (MIT core; skip ee/) plus Postgres. Helicone if that proxy is the shape. Phoenix is ELv2 — not the OSI path. No one-click template in this repo yet — Start on Railway. Serving stays on /stacks/vllm-inference.

[Start on Railway](https://c17.io/go/railway/langfuse-observe?from=%2Fstacks%2Flangfuse-observe.md)

[MLflow if this is training runs, not LLM traces](https://c17.io/stacks/mlflow-tracking)

## AGENT_PROMPT

```
Replace LangSmith with Langfuse (MIT core; skip ee/) or Helicone.

- Official compose. Persist Postgres. Point one app at the ingest URL. Do not enable ee/ and call that OSS.
- Phoenix is Elastic License 2.0 — not the OSI path. Training runs stay on /stacks/mlflow-tracking.
- Write AGENT_PROMPT.md.
```

## 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.

### What about Phoenix?

Useful, and not OSI. Elastic License 2.0 forbids offering it as a hosted service to third parties — /compare/langfuse-vs-phoenix.

### Is this Weights and Biases?

No. Training runs are /replace/wandb (MLflow). This page is LLM traces.

Source: https://c17.io/stacks/langfuse-observe