> ## Documentation Index
> Fetch the complete documentation index at: https://arizeai-433a7140.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# 07.22.2026: MCP Client Setup, Provider Filtering, and User Friction Evals
> One-command MCP client setup and a Settings guide, a playground model picker scoped to provisioned providers, a new User Friction evaluator in both SDKs, and AI SDK v7 tracing.
# One-Command MCP Client Setup
July 18, 2026
**Available in arize-phoenix 19.1.0+ (server) and @arizeai/phoenix-cli**
Following the [Remote MCP server](/docs/phoenix/integrations/remote-mcp), connecting a coding
agent to Phoenix is now a single command. `px setup mcp` writes the Phoenix MCP server into your
agent's configuration, inferring the endpoint from your active profile and defaulting to OAuth
login — no manual JSON editing.
```bash theme={null}
# Interactive: pick an agent and scope
px setup mcp
# Register the Phoenix MCP server with a specific agent
px setup mcp --agent claude
px setup mcp --agent cursor --global --endpoint https://phoenix.example.com
# Headless, with an API-key header fallback for non-interactive environments
px setup mcp --agent codex --no-input --header 'Authorization: Bearer ${PHOENIX_API_KEY}'
```
* **Supported agents** — `claude` (Claude Code), `codex`, `cursor`, `gemini`, `opencode`, and
`vscode`.
* **Global or repo-scoped** — `--global` writes user-wide config; `--local` writes repo-scoped
config in the current git repository.
* **Endpoint resolution** — the target endpoint follows the standard `px` chain (flag →
`PHOENIX_HOST` → profile → default), and `--profile` selects which profile to infer it from.
A new **MCP tab under Settings** mirrors this in the UI: it shows the deployment's MCP server URL
and whether the server and code mode are enabled, surfaces the `px setup mcp` quick-start, and
provides copy-paste connection instructions for each client. As of arize-phoenix 19.2.0, the
client list also includes **Antigravity** and **OpenCode** alongside Claude Code, Claude Desktop,
Codex, Cursor, and VS Code.
Connect Claude Code, Cursor, and other MCP clients to your Phoenix instance
# Playground Model Picker Scoped to Provisioned Providers
July 20, 2026
**Available in arize-phoenix 19.3.0+**
The playground model picker now shows only providers that are ready to use — dependencies
installed and credentials satisfied, whether stored on the server or in the browser-local
credential store — so you no longer scroll past providers you can't invoke. Custom providers you
have added appear automatically. If no provider has been provisioned yet, the picker falls back to
the flagship providers (OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, and Google) so it is never
empty.
Two Playground defaults also changed in this window: new sessions now default to OpenAI's
`gpt-5.6-sol`, and OpenAI models default to the **Responses API** (assumed whenever a prompt does
not specify otherwise).
# User Friction Evaluator
July 21, 2026
**Available in arize-phoenix-evals 3.2.0+ (Python) and @arizeai/phoenix-evals 1.2.0+ (TypeScript)**
A new built-in classification evaluator detects when a user expresses friction with an assistant's
preceding behavior — corrections, retries after an unsuccessful response, frustration, and
challenges to unrequested or unexplained actions. It returns a `friction` / `no_friction` label,
making it easy to flag conversational turns worth reviewing and to measure whether product changes
reduce expressed friction.
```python theme={null}
from phoenix.evals.metrics.user_friction import UserFrictionEvaluator
from phoenix.evals import LLM
llm = LLM(provider="openai", model="gpt-4o-mini")
user_friction_eval = UserFrictionEvaluator(llm=llm)
scores = user_friction_eval.evaluate(
{
"conversation": (
"User: Show orders from this week.\n"
"Assistant: Here are last month's orders."
),
"user_message": "No, I asked for this week.",
}
)
print(scores)
# [Score(name='user_friction', score=1.0, label='friction', ...)]
```
```typescript theme={null}
import { createUserFrictionEvaluator } from "@arizeai/phoenix-evals/llm";
import { openai } from "@ai-sdk/openai";
const evaluator = createUserFrictionEvaluator({ model: openai("gpt-4o-mini") });
const result = await evaluator.evaluate({
conversation: "User: Show recent orders.\nAssistant: Here are last month's orders.",
userMessage: "No, I asked for this week.",
});
console.log(result.label); // "friction"
```
The conversation history and the latest user message are supplied separately (`conversation` and
`user_message` / `userMessage`) so the judge classifies the target turn without confusing it with
earlier ones.
Detect expressed user friction in conversational assistants
# AI SDK v7 Tracing
July 22, 2026
**Breaking change in @arizeai/phoenix-otel 2.0.0+**
`@arizeai/phoenix-otel` now traces **Vercel AI SDK v7** applications, upgrading its
OpenInference Vercel integration to v3. Register the OpenInference span processor as before and AI
SDK v7 telemetry is translated into OpenInference spans automatically.
* **Breaking:** the upgraded integration is ESM-only and targets AI SDK v7. Applications still on
AI SDK v6 or older should stay on `@arizeai/phoenix-otel` 1.x.
* If the processor cannot be loaded (for example, a bundler strips the dynamic import), spans still
reach Phoenix through a plain OpenTelemetry processor, but AI SDK telemetry is not converted to
OpenInference.