PXI Advanced

Everything beyond the PXI quickstart: what each skill does, conversations, the terminal client, and how to configure and lock down the agent.

What it can do#

Each row is backed by a skill, a reusable procedure for one Phoenix workflow that PXI loads on demand.

File GitHub issues. When PXI finds a real defect it can file an issue linking the relevant traces, through GitHub’s MCP server. Add a fine-grained token with Issues read/write under Settings → Assistant → Tools → GitHub, or set githubPersonalAccessToken in ~/.px/settings.json for the terminal. Issues are filed as you, after approval.

Conversations#

Chats are saved on the server, so a conversation survives a reload and is the same session in the panel or the pxi client. The panel header lists past chats to continue, rename, or delete; a temporary chat is never saved. On any message, Rewind drops it and everything after, and Branch forks a new chat from that point.

Retention is set under Settings → Assistant → Chats & data; see agent session retention. Sessions are also exposed under /v1/agent_sessions; see the Agent Sessions API.

The terminal client#

The same server-side agent in your terminal, shipped with the Phoenix CLI. Credentials, skills, and permissions come from the server.

# run without installingnpx -y -p @arizeai/phoenix-cli pxi# or install it globallynpm install -g @arizeai/phoenix-cli# point it at your instance, and pick a model (defaults to claude-opus-5)export PHOENIX_ENDPOINT=http://localhost:6006   # and PHOENIX_API_KEY if auth is onpxi --provider OPENAI --model gpt-5.4

Requires Phoenix 20.0.0 or newer and a configured model provider. Lines beginning with / are handled locally: /help lists them, /compact summarizes earlier turns. See the CLI reference for every flag and command.

Configuration#

Bring your own model#

PXI runs on your model provider and your API key. Set credentials for Anthropic, OpenAI, Google, or AWS Bedrock through environment variables or Phoenix secrets, or add a custom provider under Settings → AI Providers.

Permissions and approval#

Agent assistance is opt-in.

• Turn it off per deployment (PHOENIX_DISABLE_AGENT_ASSISTANT=true), per instance (Settings → Assistant → Permissions), or per browser (Settings → Assistant → General).
• State-changing actions wait for you. Saving a prompt, writing dataset examples, or submitting an evaluator arrives as a diff or card you accept. Read-only actions run freely. Bypass Approval applies edits without asking and shows a warning while active.
• Web access is per session, through the globe button in the chat input, and only when an administrator allows it.

Privacy and tracing#

PXI runs inside your Phoenix process, on your data, with your model key. Arize is not in the request path.

• Bash runs in a sandboxed shell on the server with no network. PHOENIX_AGENTS_DISABLE_BASH=true removes it.
• PXI sees whatever you can see, so be deliberate on sensitive data.
• Chats can be recorded as traces in the assistant_agent project. Off by default; administrators turn it on under Settings → Assistant.
• PXI makes mistakes. Treat its output as suggestions.

Configuration reference#

The DISABLE and FORCE variables take true. The others take a URL, token, or name.

Feedback#

PXI is in beta. Open an issue or start a discussion on GitHub.