Skip to main content
A skill is a set of instructions your coding agent loads when a task matches it, such as instrumenting an app or writing an evaluator. Phoenix ships five, in the Phoenix repository.

Get the skills

If you connected your agent with a plugin or the MCP server, it already has the skills. There’s nothing else to install. The MCP server lists each skill’s name and a short summary in the instructions it sends when your agent connects. The agent loads a skill’s full instructions with the load_skill tool only when a task needs them. Your own custom skills reach agents the same way. phoenix-cli isn’t served over MCP because it runs px commands in a shell. To use it, install it as a file.

Install skills as files

To install skills into your agent’s skills directory, such as .claude/skills/, run:
The command asks which skills and agents to install for. To install one skill, add --skill, for example --skill phoenix-cli. Install skills as files when your agent has no MCP support, or when you use the Phoenix CLI without a plugin and want the phoenix-cli skill. An agent with both files and the MCP server sees four of the skills from both sources. That’s expected. If the two copies differ, the MCP copy is the one that matches your Phoenix server.

What your agent can do with them

cli

phoenix-cli

Debug LLM apps with the px CLI: traces, experiments, datasets, and prompts. Plugins and file installs only.
annotation

phoenix-error-analysis

Read sampled traces, write free-form notes, then group them into a failure taxonomy that picks eval targets and fix priorities.
eval

phoenix-evals

Build and run evaluators for AI/LLM apps across code-based and LLM-as-judge workflows.
trace

phoenix-tracing

Implement OpenInference tracing conventions and instrumentation in Python and TypeScript.
harbor

phoenix-harbor

Configure Harbor agent evaluations and interpret their Phoenix experiments, scores, and ATIF traces.
Example prompt, which uses phoenix-error-analysis:

Learn more

Add your own skills

Set PHOENIX_SKILLS_PATHS on the Phoenix server to one or more directories of your own Agent Skills; the MCP server serves them alongside the built-in ones:
Each skill is a directory named after the name in its SKILL.md frontmatter, which also needs a description.