Skip to main content
The fastest way to add Phoenix tracing to your application. px setup hands the instrumentation to your coding agent and doesn’t finish until a real trace arrives.
Supported stacks: Python and TypeScript/JavaScript. See the full integration list for the LLM providers and frameworks Phoenix instruments.
1

Start Phoenix

px setup needs a running Phoenix. Already have a deployment? Skip to the next step.
No uv? pip install arize-phoenix && phoenix serve does the same thing. See Terminal setup for customization.
Phoenix serves its UI and OTLP HTTP on port 6006, and OTLP gRPC on port 4317. For a local instance that’s http://localhost:6006 — leave it running while you work.
2

Connect your app

From your app’s root directory, run one of these:
px setup warns on a dirty git tree before it starts, so the agent’s edits stay separate from your own work.

​
Confirm traces are flowing

px setup verifies traces automatically. To check for yourself:
  1. Run your application and trigger at least one LLM call.
  2. Open the Phoenix UI (local: http://localhost:6006, or your deployment URL).
  3. Open the Traces view and verify traces appear under your project.
If no traces appear, check the Troubleshooting FAQ.

​
Re-run a single step

The connection questions only need answering once. On a repo that’s already registered, re-run just the slice you need:

​
Run non-interactively (CI or agents)

Pass flags instead of answering prompts:
The documentation that the MCP server offers is interactive by default. Pass --docs-mcp to connect the Phoenix docs MCP server to the coding agent without prompting, or --no-docs-mcp to skip it — either keeps a non-interactive run from stalling on a question. A run that instruments only succeeds if a trace actually arrived — the agent’s own claim that it finished doesn’t count. Exit code 6 means the wait ran out with no trace, so tracing isn’t confirmed working even though the connection, .env.phoenix, and the agent’s edits are all in place. In a pipeline, treat 6 as “configured but unverified” rather than a hard failure: re-run px setup instrument or check the exporter. In --format json|raw, the verification field carries the same verdict. See the CLI reference for the full list of flags.

​
Use an unsupported agent

px setup hands off to Claude Code, Codex, Cursor, and OpenCode. If your agent isn’t one of those (Windsurf, Copilot, and others), paste this prompt into it instead:
For ongoing agent workflows beyond initial setup (CLI, MCP, and skills), see the Coding Agents guide.