Phoenix is Arize's open-source, self-hosted AI observability platform for experimentation, evaluation, and troubleshooting of LLM and agent applications. It traces an application's runtime through OpenTelemetry-based instrumentation, scores outputs with LLM-based response and retrieval evals, keeps versioned datasets of examples, tracks experiments that compare prompt, model, and retrieval changes on the same inputs, and offers a playground for replaying traced LLM calls and comparing models. Prompt management versions and tags prompts, and PXI (Phoenix Intelligence) is an agent built into the product for debugging traces and iterating on prompts. Arize AX is the managed counterpart, and the same OpenTelemetry and OpenInference instrumentation works with both.
Phoenix is vendor, language, and framework agnostic. Instrumentation comes from the OpenInference project, with Python packages for OpenAI, OpenAI Agents SDK, Claude Agent SDK, LlamaIndex, LangChain and LangGraph, DSPy, CrewAI, Haystack, LiteLLM, Instructor, Agno, BeeAI, Pydantic AI, AutoGen AgentChat, AG2, Strands Agents, and MCP, and for providers such as Anthropic, AWS Bedrock, Google GenAI, Google ADK, Vertex AI, Mistral AI, Cohere, Groq, Together AI, and Ollama. JavaScript packages cover OpenAI, OpenAI Agents, LangChain.js, Vercel AI SDK, Mastra, BeeAI, Claude Agent SDK, MCP, TanStack AI, and Google ADK, with Java (Spring AI) and Go integrations, span processors that convert OpenLIT and OpenLLMetry traces, and platform integrations for Langflow, Flowise, Prompt Flow, LiteLLM Proxy, Envoy AI Gateway, and Graphite.
Phoenix runs on a local machine, in a container, or in the cloud, with container images and a Helm chart for Docker and Kubernetes deployments. The main Python package includes the whole platform, and lightweight Python and TypeScript sub-packages (otel, client, evals, mcp, cli) talk to a deployed instance. Code evaluators can run in hosted sandboxes (E2B, Daytona, Vercel Sandbox, Modal, Docker Sandboxes) for kernel-level isolation.
Three surfaces let coding agents such as Claude Code, Codex, and Cursor work with traces, datasets, and experiments: a CLI that fetches them and instruments an app by detecting its framework and LLM provider, agent skills shipped with the project for debugging, evaluating, and tracing with Phoenix, and a remote MCP server on each instance for querying Phoenix directly.
Features
- Tracing: OpenTelemetry-based traces of model calls, retrieval, tool use, and custom logic, viewed in real time
- Evaluation: LLM-based response and retrieval evals, plus code evaluators that can run in hosted sandboxes
- Datasets and experiments: versioned example datasets and tracked experiments that compare changes to prompts, models, and retrieval
- Playground: compares models, adjusts parameters, and replays traced LLM calls
- Prompt management: version control, tagging, and experimentation for prompt changes
- PXI: an AI engineering agent inside Phoenix for debugging traces, iterating on prompts, and navigating the product
- Remote MCP server: MCP clients query traces, datasets, and experiments on a Phoenix instance
- CLI and skills: a command-line client that fetches traces, datasets, and experiments and sets up instrumentation, plus skills for coding agents
- Instrumentation: OpenInference packages for Python, JavaScript, Java, and Go frameworks and LLM providers, with span processors for OpenLIT and OpenLLMetry traces
- Deployment: self-hosted locally, in containers, or in the cloud through Docker images and a Helm chart
- Sub-packages: otel, client, evals, mcp, and cli packages for Python and TypeScript that work with a deployed platform
