Claude Code - Arize AX Docs

Trace Claude Code CLI sessions, tool usage, and token costs with Arize AX for full observability.

Claude Code is Anthropic’s agentic coding tool that lives in your terminal. The Arize Coding Harness Tracing instruments sessions using 16 hook events and sends OpenInference spans to Arize AX. The plugin works with both the Claude Code CLI and the Claude Agent SDK.

Use this to trace Claude Code (or Agent SDK) sessions via the plugin — enabled through a settings file, no in-code instrumentor. If instead you are building an application with the Claude Agent SDK and want standard OpenInference agent, tool, and LLM spans in your app’s code, use Claude Agent SDK with ClaudeAgentSDKInstrumentor instead.

Launch Arize AX

To get started, sign up for a free Arize AX account and get your Space ID and API Key:

Log in at app.arize.com
Click Settings and copy the Space ID
Open the API Keys tab and create or copy an API key

Install

Claude Code Marketplace

claude plugin marketplace add Arize-ai/coding-harness-tracing
claude plugin install claude-code-tracing@coding-harness-tracing

The marketplace flow registers the hooks but skips the interactive wizard, so backend credentials and content-logging preferences must be set directly in ~/.claude/settings.json under env (see Configuration).

Curl installer

macOS / Linux:

curl -sSL https://raw.githubusercontent.com/Arize-ai/coding-harness-tracing/main/install.sh | bash -s -- claude

Windows (PowerShell):

iwr -useb https://raw.githubusercontent.com/Arize-ai/coding-harness-tracing/main/install.bat -OutFile $env:TEMP\install.bat
& $env:TEMP\install.bat claude

Local clone

git clone https://github.com/Arize-ai/coding-harness-tracing.git
cd coding-harness-tracing
./install.sh claude        # macOS / Linux
install.bat claude         # Windows

The curl and local installers write credentials to ~/.arize/harness/config.json and register the hooks in ~/.claude/settings.json. The installer runs a short interactive setup. Every harness in the Arize Coding Harness Tracing repo asks the same questions, in the same order.

Setup walkthrough

1. Backend selection

Choose where spans are sent:

Phoenix — your own Phoenix instance.
Arize AX — the hosted Arize platform.

2. Credentials

The prompts depend on the backend you picked.

Arize AX
Phoenix

API key — create one on the API keys tab.
Space ID — shown on the same settings tab as your API keys.
OTLP endpoint — defaults to otlp.arize.com:443. Override it only for a hosted or dedicated instance.

Endpoint — defaults to http://localhost:6006.
API key — optional. Leave it blank when Phoenix runs without auth.

If you have already configured another harness against the same backend, the installer offers a copy-from menu so you can reuse those credentials instead of retyping them.

3. Project name

The project that this harness’s spans are grouped under. Defaults to the harness name.

4. User ID (optional)

A free-form identifier attached to every span as user.id. Useful when teammates share one backend. Leave it blank to skip.

5. Content logging

Three [Y/n] opt-outs that apply to every harness, not only the one you are installing:

Log user prompts?
Log what tools were asked to do (commands, file paths, URLs)?
Log what tools returned (file contents, command output)?

You are asked these only the first time you install any harness. Later installs reuse the existing logging block in ~/.arize/harness/config.json, which you can edit at any time.

Install flags

Flag	Effect
`--with-skills`	Symlink this harness’s management skill into `.agents/skills/`, so a coding agent in the workspace can manage the tracing config for you
`--non-interactive`, `-y`	Ask nothing, and read every value from the environment instead
`--branch NAME`	Install from a specific branch instead of `main`
`--wheel-dir DIR`	Install from local wheels in `DIR`: no network access and no remote code execution

Non-interactive install

Pass --non-interactive (or -y) to skip every prompt above and take each value from the environment instead. Nothing is asked, and a missing required value is an error rather than a prompt, which makes this the mode to use from a script, from CI, or when a coding agent is driving the install itself. Values come from the environment, or from a dotenv file named explicitly with ARIZE_ENV_FILE. Naming a file keeps the API key out of the command line and your shell history.

ARIZE_ENV_FILE=~/.arize/onboarding.env ./install.sh <harness> --non-interactive

Variable	Default	Description
`ARIZE_API_KEY` + `ARIZE_SPACE_ID`	—	Arize AX credentials. Both are required for the Arize backend.
`PHOENIX_ENDPOINT`, `PHOENIX_API_KEY`	`http://localhost:6006`	Phoenix endpoint and optional API key.
`ARIZE_BACKEND`	inferred	`arize` or `phoenix`. A space ID implies Arize AX and a Phoenix endpoint implies Phoenix. When both are present the install stops and asks you to set this rather than guess, since guessing would discard one backend’s credentials.
`ARIZE_PROJECT_NAME`	harness name	Project that spans are grouped under. Read from the dotenv file only. An installed harness exports its own project name into every session, so an inherited environment value is ignored here.
`ARIZE_USER_ID`	—	Optional `user.id` on every span.
`ARIZE_OTLP_ENDPOINT`	`otlp.arize.com:443`	Override for a hosted or dedicated Arize instance.
`ARIZE_LOG_PROMPTS`	`false`	Set `true` to capture prompt text.
`ARIZE_LOG_TOOL_DETAILS`	`false`	Set `true` to capture tool commands, file paths, and URLs.
`ARIZE_LOG_TOOL_CONTENT`	`false`	Set `true` to capture tool output.
`ARIZE_ENV_FILE`	—	Dotenv file to read. No file is read unless this is set, and a path that is not a readable file is an error rather than a fallback to the environment.
`ARIZE_WHEEL_DIR`	—	Same as `--wheel-dir`. Install from local wheels instead of downloading the repo.

A named ARIZE_ENV_FILE outranks the environment, and there is deliberately no automatic ./.env search. Reading the working directory would let a cloned repository’s dotenv choose ARIZE_OTLP_ENDPOINT or PHOENIX_ENDPOINT while your real credentials came from the environment, installing a config that ships spans and a bearer API key to an endpoint the repo picked, for every later session on that machine. Name the file you mean.

Content logging is off by default in this mode, unlike the interactive wizard where each question defaults to yes. A [Y/n] default is a person declining to change an answer they were shown; the same default unattended would capture prompts, commands, and file contents that nobody agreed to. Set the ARIZE_LOG_* variables you want to true.

The API key is never echoed. The installer reports only that it found one and where it came from, and every resolved value is reported with its source, so a wrong-credentials install stays diagnosable:

[arize] Backend: Arize AX at otlp.arize.com:443 (from default)
[arize]   space ID: my-space (from /path/to/.env)
[arize]   API key: found (from /path/to/.env)
[arize] Project name: codex (from default)

Check what’s installed

status reports which harnesses are configured and whether their hooks are actually wired into each harness’s own settings file. Both have to be true for traces to appear.

./install.sh status
./install.sh status --json    # machine-readable

hooks: NOT registered means credentials are saved but the harness was never wired up, or something removed the hooks. Re-run the install for that harness. Use --json from a script or a coding agent to gate on the exit code without parsing output: 0 means every configured harness is wired up, 1 means nothing is configured, and 2 means at least one harness’s hooks are missing. The payload contains no secrets — an API key appears only as "api_key_present": true — so it is safe to paste into a bug report.

Keep it up to date

update pulls the latest code and re-registers every harness already in config.json.

./install.sh update

Re-registering runs each harness’s installer, so in a terminal it still asks for each project name. With no terminal to answer on, in CI or a cron job, it takes the stored values instead of failing: credentials are not re-read on that path, and the project name keeps whatever is in config.json.

Configuration

The curl and local installers write credentials to ~/.arize/harness/config.json. Environment variables in ~/.claude/settings.json take precedence and are required for the marketplace install path.

{
  "env": {
    "ARIZE_API_KEY": "<your-api-key>",
    "ARIZE_SPACE_ID": "<your-space-id>",
    "ARIZE_PROJECT_NAME": "claude-code",
    "ARIZE_TRACE_ENABLED": "true"
  }
}

Redaction controls

Each ARIZE_LOG_* flag accepts "true" or "false" and defaults to "true". Set to "false" to opt out per category:

{
  "env": {
    "ARIZE_LOG_PROMPTS": "false",
    "ARIZE_LOG_TOOL_DETAILS": "false",
    "ARIZE_LOG_TOOL_CONTENT": "false"
  }
}

Flag	Redacts
`ARIZE_LOG_PROMPTS`	User prompt and assistant response text
`ARIZE_LOG_TOOL_DETAILS`	Tool names and arguments
`ARIZE_LOG_TOOL_CONTENT`	Tool call output content

Observe

Now that you have tracing set up, all Claude Code sessions stream to your Arize AX account for observability and evaluation. You’ll see:

Turn traces — each conversation turn (user prompt → assistant response)
LLM spans — Claude’s responses with model info and token counts
Tool spans — nested spans for each tool call with inputs, outputs, and duration
Subagent spans — activity from any subagents Claude spawns
Session grouping — all turns from the same session grouped by session_id

Claude Code session view in Arize AX showing multiple traces grouped by session ID — Claude Code turns grouped together in a single session view

Drill into any turn trace to inspect the full span tree, including model generations, tool calls, and subagent activity.

Claude Code trace view in Arize AX showing the trace tree, inputs, outputs, and nested spans — Detailed trace view for a Claude Code turn

Hooks Captured

Hook	Span Kind	Description
`SessionStart`	CHAIN	Session initialized, trace/tool counters reset
`UserPromptSubmit`	CHAIN	User prompt captured (also lazy-inits session for SDK)
`UserPromptExpansion`	—	Stashes slash command metadata to attach to the next turn span
`PreToolUse`	TOOL	Tool invocation started, records tool name and input
`PostToolUse`	TOOL	Tool invocation completed, records output and duration
`PostToolUseFailure`	TOOL	Tool invocation failed, records error attributes
`Stop`	LLM	Model response completed with input/output values
`StopFailure`	LLM	Turn failed before completion, emits a span with `error.type` / `error.message`
`SubagentStart`	—	Records subagent start time and prompt keyed by `agent_id`
`SubagentStop`	CHAIN	Subagent response completed
`PreCompact`	—	Records compaction start time and trigger
`PostCompact`	CHAIN	Compaction completed, emits a span when it fires inside a turn
`PermissionRequest`	CHAIN	Permission prompt for tool use
`PermissionDenied`	CHAIN	Auto-mode tool denial recorded inside the current turn
`Notification`	CHAIN	System notification event
`SessionEnd`	CHAIN	Session teardown, state file cleanup

Hooks marked — don’t emit their own span — they stash state that an adjacent hook (Turn, tool, or compaction) attaches when it fires.

Agent SDK Setup

The tracing plugin also works with the Claude Agent SDK in both Python and TypeScript. The SDK loads the plugin locally — no marketplace install is required — but the setup must be done in your application code before the SDK session starts, so the agent cannot configure it at runtime.

You must use ClaudeSDKClient. The standalone query() function does not support hooks, so tracing will not work with it.

1. Locate the plugin

The plugin path depends on how you installed the harness:

Installed via the Claude Code CLI marketplace: the plugin is cached at ~/.claude/plugins/cache/coding-harness-tracing/claude-code-tracing/1.0.0.
Installed via the curl or local installer: the plugin lives at ~/.arize/harness/tracing/claude_code.
Not installed: clone the repo into your project — the plugin path is ./coding-harness-tracing/claude-code-tracing:
git clone https://github.com/Arize-ai/coding-harness-tracing.git

2. Create a settings file

The SDK spawns a Claude Code subprocess that does not inherit your shell environment, so tracing env vars must be passed through a settings file referenced from ClaudeAgentOptions:

{
  "env": {
    "ARIZE_TRACE_ENABLED": "true",
    "ARIZE_API_KEY": "<your-api-key>",
    "ARIZE_SPACE_ID": "<your-space-id>",
    "ARIZE_PROJECT_NAME": "claude-code"
  }
}

The same ARIZE_LOG_* redaction flags from Configuration apply here.

3. Wire the plugin into your app

Pass the plugin path and settings file to ClaudeSDKClient:

Python
TypeScript

from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient

PLUGIN_PATH = "./coding-harness-tracing/claude-code-tracing"  # or your install path

options = ClaudeAgentOptions(
    plugins=[{"type": "local", "path": PLUGIN_PATH}],
    settings="./settings.local.json",
)

async with ClaudeSDKClient(options=options) as client:
    await client.query("Your prompt here")
    async for message in client.receive_response():
        print(message)

import { ClaudeSDKClient } from "@anthropic-ai/claude-agent-sdk";

const PLUGIN_PATH = "./coding-harness-tracing/claude-code-tracing"; // or your install path

const client = new ClaudeSDKClient({
  plugins: [{ type: "local", path: PLUGIN_PATH }],
  settings: "./settings.local.json",
});

await client.connect();
await client.query("Your prompt here");
for await (const message of client.receiveResponse()) {
  console.log(message);
}
await client.close();

If you installed via the curl or local installer, the harness ships a Python convenience helper that returns a pre-configured ClaudeAgentOptions (plugin path + setting_sources=["user"] so user-level Claude settings are honored):

from tracing.claude_code.agent_sdk import claude_options

async with ClaudeSDKClient(options=claude_options()) as client:
    ...

Validate

Add "ARIZE_DRY_RUN": "true" to your settings file to verify hooks fire without sending data, and tail ~/.arize/harness/logs/claude-code.log to confirm activity.

Hook parity

SDK	Coverage
TypeScript	Full parity — all 16 hooks fire, including `SessionStart`, `Notification`, `PermissionRequest`, and `SessionEnd`.
Python	`SessionStart`, `SessionEnd`, `Notification`, and `PermissionRequest` are not available. Session state is lazily initialized on the first `UserPromptSubmit`; core tracing (LLM, tool, and subagent spans) still works fully.

Reference

For the full list of environment variables, default file paths, and troubleshooting steps, see the Claude Code tracing README.

Uninstall

Marketplace install:

claude plugin uninstall claude-code-tracing@coding-harness-tracing
claude plugin marketplace remove Arize-ai/coding-harness-tracing

Curl or local install:

curl -sSL https://raw.githubusercontent.com/Arize-ai/coding-harness-tracing/main/install.sh | bash -s -- uninstall claude

​Launch Arize AX

​Install

​Claude Code Marketplace

​Curl installer

​Local clone

​Setup walkthrough

​1. Backend selection

​2. Credentials

​3. Project name

​4. User ID (optional)

​5. Content logging

​Install flags

​Non-interactive install

​Check what’s installed

​Keep it up to date

​Configuration

​Redaction controls

​Observe

​Hooks Captured

​Agent SDK Setup

​1. Locate the plugin

​2. Create a settings file

​3. Wire the plugin into your app

​Validate

​Hook parity

​Reference

​Uninstall

​Resources

Arize Coding Harness Tracing

OpenInference

Claude Code Documentation

Claude Code Plugins

CLAUDE.md: Best Practices for Optimizing Claude Code with Prompt Learning

Launch Arize AX

Install

Claude Code Marketplace

Curl installer

Local clone

Setup walkthrough

1. Backend selection

2. Credentials

3. Project name

4. User ID (optional)

5. Content logging

Install flags

Non-interactive install

Check what’s installed

Keep it up to date

Configuration

Redaction controls

Observe

Hooks Captured

Agent SDK Setup

1. Locate the plugin

2. Create a settings file

3. Wire the plugin into your app

Validate

Hook parity

Reference

Uninstall

Resources