Arize Phoenix

## Key capabilities

- **Tracing** — captures LLM/agent spans (prompts, tool calls, retrievals) using OpenTelemetry, so traces are portable across instrumented frameworks.
- **Evaluation** — run LLM-as-a-judge and heuristic evals over traces and datasets to score relevance, hallucination, and task success.
- **Datasets & experiments** — curate examples from production traces and compare prompt/model versions side by side.
- **Self-hosted or local** — run Phoenix locally or in your own infrastructure, keeping trace data in your environment.

## How Phoenix compares

Phoenix sits in the LLM observability/evaluation space alongside several tools also in this directory. Key differences:

| Tool | Type | Self-hostable | Notable for |
| --- | --- | --- | --- |
| **Arize Phoenix** | Open-source observability + evals | Yes | OpenTelemetry-based tracing that runs locally |
| **LangSmith** | Proprietary observability + evals | Enterprise tier | Deep LangChain / LangGraph integration |
| **Helicone** | Observability via a request-logging proxy | Yes | One-line integration that captures requests |
| **Braintrust** | Proprietary evals + experimentation | SaaS | Eval-first experimentation workflow |

Choose Phoenix when you want open-source, OpenTelemetry-native tracing and evaluation you can run locally; pair it with Helicone if you also need request-level logging.

## Editorial notes

Phoenix is useful for teams that want open-source tracing and evaluation workflows around agent and LLM behavior.

## Disclosure

Editorial listing. No paid placement or affiliate link is used.