​

Phoenix Plugins for Coding Tools

claude plugin marketplace add Arize-ai/phoenix
claude plugin install arize-phoenix@arize-phoenix

codex plugin marketplace add Arize-ai/phoenix
codex plugin add arize-phoenix@arize-phoenix

• Claude Code and Cursor also get the skills. They install phoenix-cli together with phoenix-evals and phoenix-tracing, so there is no separate skills add step. The Codex plugin registers the MCP server config.
• Point it at your instance once. Claude Code and Cursor prompt for a Phoenix endpoint and default to http://localhost:6006. They append /mcp themselves; Codex reads PHOENIX_ENDPOINT from the shell you launch it in.
• OAuth handles auth. When your Phoenix requires a login, the MCP server signs in through the browser on first use. For a headless environment, register a bearer-token server with px setup mcp instead.
• Codex can use an API key without exposing it. If PHOENIX_API_KEY is set, the launcher passes it to mcp-remote at runtime instead of putting the plaintext key in the command arguments.
• Cursor also gets Phoenix docs. The Cursor plugin registers a phoenix-docs MCP server next to the instance server, so it can search the docs without extra setup.
• Manual setup remains available. You can still install the CLI and MCP server separately, then add skills only when you need them.

​

Error Analysis Through MCP

• Any MCP client can load it because the /mcp handshake lists the shared skills and exposes load_skill and load_skill_reference, so a connected tool can run it without a local skill install.
• PXI records notes on the trace data. PXI can create notes on spans or on whole traces and sessions through MCP, so open coding notes land on the entity instead of in the chat log. This replaces the older debug-trace / span-coding / annotate-spans skills.
• Summary links come prefiltered. Each annotated level links to its own filtered traces or sessions table.
• Install it anywhere with npx skills add Arize-ai/phoenix --skill phoenix-error-analysis.

​

Completeness Evaluator

• Completion means finished work. Delivered answers and artifacts count only when they include the required parts. Actions count only when success is visible in the record. Refusals do not count as completed work; neither do clarifying questions or blocker reports. Withdrawn requests are excluded.
• Pass the whole record as conversation. For traced conversations, include tool results next to the calls so the evaluator can verify that an action actually succeeded.
• The result is one label. It returns complete or incomplete, with an explanation that walks each request it tracked.

from phoenix.evals import LLM
from phoenix.evals.metrics import CompletenessEvaluator

evaluator = CompletenessEvaluator(llm=LLM(provider="openai", model="gpt-4o-mini"))

scores = evaluator.evaluate(
    {
        "conversation": (
            "User: Reset my password and update the billing address.\n"
            "Assistant: Your password has been reset."
        )
    }
)
print(scores[0].label)  # "incomplete"

import { createCompletenessEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

const evaluator = createCompletenessEvaluator({ model: openai("gpt-4o-mini") });

const result = await evaluator.evaluate({
  conversation:
    "User: Reset my password and update the billing address.\nAssistant: Your password has been reset.",
});
console.log(result.label); // "incomplete"

​

Document Relevance Evaluators Are Deprecated

• Rename the input field document_text to context; in TypeScript, the field is documentText.
• The negative label changes from unrelated to irrelevant, so update anything that branches on it.
• Pass one document as context to keep scoring each document separately.

from phoenix.evals import LLM
from phoenix.evals.metrics import RetrievalRelevanceEvaluator

evaluator = RetrievalRelevanceEvaluator(llm=LLM(provider="openai", model="gpt-4o-mini"))

scores = evaluator.evaluate(
    {
        "input": "What is the capital of France?",
        "context": "Paris is the capital and largest city of France.",
    }
)

​

Filter Spans by Cost and Annotation Identifier

• Cost scalars total_cost / prompt_cost / completion_cost read the span’s cost row and return 0 when the span has no recorded cost.
• cost_details iterates the rows for each token type with any, all, len, sum, max, and min, exposing token_type, is_prompt, cost, tokens, plus cost_per_token.
• .identifier joins .label / .score / .explanation on annotation lookups in the span, trace, or session filters.
• The filter bar autocompletes the new names and includes snippets for cost and cost-detail filters.

total_cost > 0.10
any(cost_detail.token_type == 'cache_read' for cost_detail in cost_details)
annotations['quality'].identifier == 'pxi'

​

Prompt Version Metadata

• Set and read it over REST or GraphQL, then pass metadata to the TypeScript promptVersion() helper.
• The playground’s save dialog offers the Metadata field when saving a new version of an existing prompt, not only when creating a prompt.
• The prompt version details page displays it, so you can inspect version tags without writing a query.

import { createPrompt, promptVersion } from "@arizeai/phoenix-client/prompts";

await createPrompt({
  name: "support-triage",
  version: promptVersion({
    modelProvider: "OPENAI",
    modelName: "gpt-4o-mini",
    metadata: { owner: "support-eng", reviewed: true },
    template: [{ role: "user", content: "Classify this ticket: {{ticket}}" }],
  }),
});

​

REST API Updates

• filter on GET /v1/projects/{project_identifier}/traces and .../sessions takes the same trace and session filter expressions the UI uses. It combines with the other query parameters using AND; an empty expression does not filter, and an invalid one returns 400.
• The discrete trace filters are deprecated. error still works, as do min_latency_ms and max_latency_ms, but filter=error_count > 0 and filter=latency_ms >= 1000 replace them.
• GET /v1/datasets/{dataset_identifier}/splits lists a dataset’s splits with cursor pagination. A split appears when at least one of its examples belongs to the dataset, so example_count is dataset-scoped and agrees with the create/update/delete endpoints.

curl -G "$PHOENIX_ENDPOINT/v1/projects/my-project/traces" \
  --data-urlencode "filter=error_count > 0 and total_cost > 0.05" \
  -H "Authorization: Bearer $PHOENIX_API_KEY"

​

Additional Improvements

• Reasoning content appears in its own collapsible block for LLM span messages, with Markdown summaries and an explanation when the provider returned only an encrypted payload.
• Token tooltips show cache reads and writes, so cumulative counts in the traces, spans, session, and experiment views show whether prompt caching was hit.
• Trace and session filters live in the URL. traceFilterCondition and sessionFilterCondition make a filtered table shareable, and a filtered link no longer flashes unfiltered rows first.
• The traces table shows span annotations by default, under a column name that accounts for the child spans revealed by expanded rows.
• Experiment comparisons open example details from an ID and show example external IDs.
• Notes have GraphQL mutations at every supported record level, with an explicit annotator kind and source.
• Built-in token prices are refreshed, image generation models stay in the cost manifest, and gpt-image-2.5 is priced.
• Annotation explanation controls are clickable and keyboard-reachable, and code editor errors render inside dialogs instead of overflowing them.
• The bundled SQLite extension driver fixes memory-safety and correctness bugs in arize-phoenix-sqlean 0.1.2.