Phoenix

Arize built Phoenix as a self-hosted workbench for debugging LLM and agent apps through their traces. Instrumentation comes from the companion OpenInference project, OpenTelemetry-based and Apache-2.0, with auto-instrumentors for OpenAI, Anthropic, LangGraph, LlamaIndex, CrewAI and many others across Python, JavaScript, Java and Go. Collected spans land in a web UI where you inspect tool calls and retrieval steps, attach LLM-judged or code-based evaluations, turn traces into versioned datasets, compare prompts or models in experiments, replay calls in a playground and keep prompt versions; a remote MCP endpoint exposes the same data to coding agents. To try it, pip install arize-phoenix and run phoenix serve; production setups use the Docker image or Helm chart, with SQLite by default and PostgreSQL 14+ for multi-user installs. Its license, Elastic 2.0 (ELv2), lets companies run and modify it internally at no cost but forbids reselling it as a hosted service, so it is source-available rather than OSI open source. UI analytics are on by default but optional. Suits teams wanting a local alternative to hosted tracing tools.