Arize Phoenix
App : A product you use directly
LLM application tracing and evaluation, built on OpenTelemetry.
Excellent locally for debugging an agent; Elastic license, not OSI.
Security reflects where data goes, requested permissions and incidents on record at the verification date. It is not a guarantee of zero risk: check the deployment mode and the connections you enable.
What it does, concretely
When building an application on top of a language model, it is hard to understand why a response is wrong: is it the prompt, the document retrieval, a tool call, or the model itself? Arize Phoenix solves this by recording every execution step of the application as traces, making it possible to pinpoint exactly where something went wrong.
Phoenix works by instrumenting your code via OpenTelemetry and the OpenInference specification. It collects traces from your LLM calls, displays them in a web interface, and lets you automatically evaluate response quality using an LLM as a judge. You can also create datasets from your traces, run experiments to compare prompt variants, and manage prompts with a versioning system.
The tool is aimed at developers and AI engineers building agents or RAG pipelines who need visibility into their behavior in development or production. It is compatible with major frameworks (LangGraph, LlamaIndex, CrewAI, OpenAI Agents SDK, etc.) and model providers (OpenAI, Anthropic, Google, AWS Bedrock, etc.), with no lock-in to a single ecosystem.
Phoenix is not suited for teams looking for a fully managed solution with no infrastructure to maintain: in that case, Arize offers Arize AX, a separate cloud platform. The Elastic license (ELv2) also prohibits offering Phoenix as a commercial service to third parties.
What it offers
- ·OpenTelemetry tracing: every LLM call, document retrieval, and tool call is recorded and visualized in a web interface.
- ·Automatic evaluation: an LLM is used as a judge to score relevance, quality, and other criteria on your traces.
- ·Dataset management: creation of versioned datasets from traces, usable for experiments or fine-tuning.
- ·Experiments: comparison of prompt variants, models, or retrieval parameters on the same dataset.
- ·Prompt playground: test and optimize prompts directly in the interface, with the ability to replay traced calls.
- ·Prompt management: versioning, tagging, and systematic testing of prompts through a dedicated interface.
- ·PXI (Phoenix Intelligence): an AI agent built into Phoenix for debugging traces, iterating on prompts, and navigating the product.
- ·Remote MCP server: connect MCP clients (Claude Code, Cursor, etc.) directly to your Phoenix instance to query traces and datasets.
In practice
- Price
- Freemium
- Commercial use
- Commercial use: yes
- Available on
- web, docker, cli
- Licence
- Elastic-2.0
Security and data
- Rating
- Security: nothing to flag
- Your data
- Data stays local
- Known incidents
- Rachat d'Arize par Dynatrace annoncé le 13 août 2026.
- Maintenance
- Active
- Last release
- 24 October 2026
Links and repository checked on 25 October 2026 · added on 1 April 2023
How to install Arize Phoenix
pip install arize-phoenix