How do I get started with Phoenix for free?

You can run Phoenix locally with a single command: `uvx arize-phoenix serve`. It starts a server at localhost:6006 with no account required.

What does the 'Signal' feature in pricing refer to?

Signal refers to the number of AI issues (like errors or problematic outputs) you can track and alert on per month.

Does Phoenix support non-Python languages?

Yes, Phoenix has an official TypeScript SDK and uses OpenTelemetry, making it compatible with any language that supports OTEL instrumentation.

Can I use my own LLM provider for evaluations in the free tier?

No, the free and Pro tiers use Arize-managed AI providers for evaluations. You can bring your own key only in the Enterprise tier.

How does Phoenix handle data privacy and compliance?

Phoenix can be fully self-hosted to keep data on your own infrastructure. Enterprise plans include SOC 2 Type II, GDPR, and HIPAA compliance options.

What are 'Custom Managed Agents'?

These are proprietary Arize agents like Alyx that are managed in the cloud. Access to them and their repository integrations is limited to Enterprise plans.

Directory/Analytics/AI Development/Arize Phoenix

Arize Phoenix

Arize Phoenix is an open-source, local-first platform designed for tracing, evaluation, experimentation, and prompt iteration of LLM-based applications and AI agents. It provides full visibility into agent trajectories, allowing you to debug issues by tracing every step from prompts and retrievals to tool calls and outputs. Key capabilities include LLM-as-a-judge evaluations, dataset and experiment management, and a prompt IDE for testing changes with evidence. It's built for AI engineers and developers who need a systematic way to observe, annotate, hypothesize, experiment, and measure the quality of their AI systems.

Analytics

Visit website

Added on

Sep 22, 2026

Open-sourceAI observabilityAgent tracingLLM evaluationPrompt engineeringOpenTelemetrySelf-hosted

Product information

Everything worth knowing about Arize Phoenix

The complete picture — from everyday features to the technical detail builders and IT folks dig for.

Core features

What it actually does

Agent Observability & Tracing

Trace every step of your LLM agent: prompts, retrievals, tool calls, and outputs. Understand why an agent responded a certain way.

LLM & Agent Evaluation

Build evaluations (evals) that score outputs for quality, relevance, toxicity, and more using built-in checks or LLM-as-a-judge.

Datasets & Experimentation

Create datasets from traces, run experiments to test changes, and use the Prompt IDE for iterative development.

Human Annotation & Feedback

Mark what worked, flag what broke. Add labels with human review, manage labeling queues, and collect feedback.

Technical capabilities

OpenTelemetry Compliant

Built on the OpenTelemetry standard for traces and instrumentation, avoiding vendor lock-in and enabling use with existing tools.

Multi-Modal & Agent-Native Tracing

Supports tracing for complex agent architectures, including multi-turn conversations and multi-modal data (images, voice, PDF).

Must watch videos

Tutorial

How to Evaluate AI Agents: Full Course

AI Agents Mastery is a popular hands-on course we are updating for 2026 for AI engineers and ambitious builders who want to go beyond toy chatbots and ship real, production-ready agents. Over seven modules and a series of guided labs, you’ll learn how to design agent workflows, wire up tools and the Model Context Protocol (MCP), build RAG and agentic RAG, add deep observability, and evaluate your agents with robust metrics and LLM-as-a-judge patterns. You’ll work directly in Python with a GitHub lab repo, instrumenting and improving a single agent system as you go—from the very first “hello world” request all the way to post-deployment monitoring, cost tracking, and guardrails—so you leave with both conceptual understanding and a working, extensible agent stack.

Tutorial

Arize Observe 2026

The AI Agent Evals Conference | June 4, 2026 in the San Francisco Ferry Building

Tutorial

Rise of the AI Engineer

If LLMs were the spark, AI and agent engineers are the ones building the power grid. Blending a unique set of skills – part SWE, part systems/infra, part AI researcher – they are the heroes who actually make agents work in the wild. This series spotlights agent engineering thought leaders to dissect their use cases, best practices, and what they see on the bleeding edge.

Who uses it

Real use cases, no hype

LLM Agent Debugging

Trace every step of your agent's decision-making process to pinpoint failures in prompts, retrieval, or tool calls.

Production AI Monitoring

Monitor token usage, latency, and costs in real-time to ensure AI performance and manage budgets.

Prompt Engineering & Iteration

Use the Prompt IDE to test, compare, and version prompts, then run experiments with trace replay to validate improvements.

Automated Evaluation (LLM-as-a-Judge)

Build and run offline/online evaluations to score model outputs for quality, relevance, and safety before deployment.

Multi-turn Conversation Analysis

Session support allows you to track and evaluate complete user interactions across multiple LLM calls.

Multi-modal AI Development

Trace and evaluate applications that process images, voice, and documents alongside text-based LLMs.

A/B Testing for Model Changes

Create datasets from traces and run controlled experiments to measure the impact of model or prompt changes on performance.

What's great

Full platform for tracing, evaluation, and prompt engineering.
Generous free tier: 25k trace spans, unlimited users & evals.
Open-source core (Phoenix) with a clear path to managed cloud.

Technical strengths

Native OpenTelemetry compliance for vendor-agnostic instrumentation.
Multi-modal tracing support for images, voice, and PDFs.
Self-hosted deployment option for data privacy and control.

Where it falls short

Free tier is limited: 10 issues/month, 1 GB storage, 15-day retention.
Advanced agent features (repo access, swarm observability) are Enterprise-only.
Pro plan price jumps to $50/month for modest limits increase.

Technical limitations

Custom Managed Agents and advanced security features are not available in lower tiers.
Some powerful evaluation tools (custom code evaluators, align evaluators) require Enterprise.
Pricing for trace ingestion can scale quickly with high-volume production use.

For technical folks

The deep-dive specs

FAQ

Includes technical Q&A

Phoenix is the open-source, local-first platform for tracing and evaluation. Arize AX is the managed SaaS version that adds features like custom managed agents, SSO, and enterprise support.