tools

Arize Phoenix

Piattaforma open source per tracciare, valutare e migliorare gli agenti di IA e le applicazioni LLM.

In inglese

Arize Phoenix traces prompts, retrieval, tool calls, outputs, latency, and token usage. It supports evaluations, human annotations, datasets, experiments, prompt iteration, and OpenTelemetry-based integrations.

Developers and AI engineering teams can run it locally, in Docker, Kubernetes, notebooks, or Phoenix Cloud. Phoenix is self-hosted and open source; managed Arize AX plans and dedicated support are separate offerings.

Features

Trace prompts, retrievals, tool calls, outputs, and agent steps
Evaluate AI outputs with code evaluators or LLM-as-judge
Create datasets from traces and run repeatable experiments
Annotate traces with human feedback or automated labels
Compare and manage prompt versions in the Prompt IDE
Support OpenTelemetry and OpenInference tracing standards
Run locally, in Docker, Kubernetes, notebooks, or Phoenix Cloud
ELv2-licensed open-source core

Use cases

Debug failed agent runs by inspecting traces and tool calls
Evaluate retrieval quality and generated responses before release
Create benchmark datasets from production traces
Compare prompt or model changes through controlled experiments
Monitor token usage, latency, errors, and AI application behavior

Pros

Cons

Latest updates

2026-10-29 Get a webhook when an eval task fails
Subscribe a webhook to the new ONLINE_TASK_RUN_FAILED event, so you hear when an eval task stops writing evals.
2026-10-28 Fixes and improvements
Fix experiments SDK warnings, proxy authentication, Vertex AI Gemini routing, dashboard filters, custom metrics, annotation status, and Alyx.
2026-10-23 Run managed agents on the OpenCode harness
OpenCode joins Claude Code, Codex, and Cursor as a harness for managed agents, and is available on every account.
2026-10-21 Find and fix eval gaps with the Eval Health agent
Reviews evals on a tracing project and proposes changes to create, fix, delete, or optimize evaluators for approval.
2026-10-21 Score whole conversations with trace and session remote evaluators
Remote evaluators at trace and session scope receive mapped variables as arrays, with one value per span, to score a conversation in one call.

Capabilities

Command line — “uvx arize-phoenix serve” source
Self-hosted — “Run locally or self-host in your environment.” source
API — “Quickstarts, API reference, and integration guides” source
Traces and evaluates — “Trace, evaluate, and improve AI agents with Phoenix.” source

Get it

Documentation

Security

SOC 2 Type II — “Arize AI Announces SOC 2 Type II Certification” source
SOC 2 — “Arize AI Announces SOC 2 Type II Certification” source
ISO 27001 — “ISO/IEC 27001 Certified” source
GDPR — “GDPR Compliant” source
HIPAA — “HIPAA Compliant” source
Data kept in the EU — “Arize AI: Support for EU Data Residency” source

Pricing

Starting price: $50/mo
Prices checked: 2026-10-26

AX Free

Free

10 issues per month
25k spans per month
1 GB per month
15 days retention
SaaS deployment
Unlimited users and evals

AX Pro

$50 per month

25 issues per month
50k spans per month
10 GB per month
30 days retention
SaaS deployment
Unlimited users and evals

AX Enterprise

Price on request

Unlimited signal issues
Custom span volume
Custom ingestion volume
Custom retention
SaaS or self-hosted deployment
Unlimited users and evals

Official website