Get started
1
Configure a model
PXI needs a model to talk to. Set credentials for at least one provider
(
OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, AWS Bedrock
credentials, or a custom provider under Settings → Models). See
Setup for the validated model list.2
Open PXI and accept the consent gate
Open the assistant from any Phoenix page — click the assistant button or
press ⌘I (Ctrl+I on Windows and Linux). The first time,
you review the session-trace settings and acknowledge to enable the chat
surface for your browser.
3
Ask it something
From a failing trace, try “why did this fail?” From the prompt playground,
try “make this prompt more robust to empty input.” PXI uses the page you
are on as context. No need to paste IDs or copy data.
Conversations
Your chats with PXI are saved on the Phoenix server, not in the browser tab. A conversation survives a reload, follows you to another browser, and is the same session whether you opened it in the in-browser panel or thepxi terminal
client. The chat panel keeps a session list in its header: pick a past chat to
continue it, rename one (Phoenix seeds a title from the opening turn), or delete
the ones you no longer need. Older sessions load as you scroll.
Temporary chats are the exception. A chat you mark temporary is never written to
your history and carries an ephemerality badge. Toggle it per chat before you
send the first message, or make it the default under Settings → Assistant →
Personal settings → Start new chats as temporary.
Rewind and branch. Every message carries a control to change course from that
point:
- Rewind drops the message and everything after it. On one of your own messages the text drops back into the input so you can edit and re-send it; on a response the chat reverts to that answer and discards what came after.
- Branch forks a new chat from that point and leaves the current one untouched — a way to try an alternative without losing the thread you came from.
Over the API
The same sessions are exposed as REST routes under/v1/agent_sessions — create,
list, get, patch, compact, chat, submit tool outputs, and fetch messages — the
endpoints the browser panel and the terminal client both use. See the
Agent Sessions API reference.
The TypeScript client @arizeai/phoenix-client exports an AGENT_SESSION_*
capability requirement for each route so a program can check server support
before calling.
Rewind, branch, and delete are GraphQL-only. The REST routes cover create,
chat, and compaction, but the operations behind the rewind and branch controls
above have no /v1/agent_sessions equivalent. The browser panel performs them
through the truncateAgentSession (rewind), branchAgentSession (branch), and
deleteAgentSession mutations on the Phoenix GraphQL API. A client that needs to
reproduce that UX has to call GraphQL.
One turn runs against a session at a time. A streaming turn holds a lock that it
refreshes with a periodic heartbeat; a lock whose heartbeat goes stale (roughly a
minute without a refresh) can be reclaimed by the next turn. A request that
collides with another client comes back as 409 Conflict with a code that says
what to do:
Use PXI from the terminal
PXI is also available as an interactive terminal chat, shipped with the Phoenix CLI (@arizeai/phoenix-cli). It is the same server-side agent that powers the
in-browser experience. The CLI connects to a running Phoenix instance, so model
credentials, skills, and permissions are configured on the server exactly as
described under Setup.
Run it without installing:
pxi command directly:
PHOENIX_ENDPOINT (and PHOENIX_API_KEY
if your deployment requires auth) environment variables, or pass --endpoint /
--api-key flags:
--provider and --model (defaults to Anthropic
claude-opus-5):
The terminal client talks to the same
/v1/agent_sessions endpoints as the
browser agent (sending headless: true on its chat turns), so it
requires a running Phoenix server with a configured model provider.
On launch it runs a preflight check against the server’s model catalog and
credentials, surfacing configuration problems as a clean error before the chat
opens. It also requires a Phoenix server on 20.0.0 or newer — the release
that introduced persisted agent sessions — and exits at startup with an upgrade
message when the connected server is older, rather than failing on the first
send.
Each capability flag is still subject to the server’s own settings — for
example,
--enable-subagents has no effect when
PHOENIX_AGENTS_DISABLE_BASH=true, and --export-remote-traces requires a
configured collector and an administrator who has allowed export.
Slash commands
Inside the terminal chat, lines beginning with/ are handled locally by the
client and never sent to the model.
/compact asks the model to summarize every completed turn into a durable
checkpoint; later turns load history from the checkpoint onward, freeing
context for long-running sessions. Compaction is rejected while the session is
in use elsewhere (for example, a turn streaming in the browser). The chat
shows the busy indicator and refreshes when the other turn completes.
Status line
While a turn streams, the terminal client shows a live thinking indicator, and once the response settles a bottom-right context-usage line reports the size of the context PXI is currently carrying — the tokens retained after the latest response, not a running total for the turn — preceded by a cache-activity summary when the provider reports cache reads or writes. The active model name sits next to the input prompt so you always know which provider and model the session is talking to. In the browser the same usage appears under the chat as a compact total you can expand into a prompt-versus-completion breakdown. When the provider caches part of the prompt, the prompt figure splits further into uncached, cache-read, and cache-write tokens.What it does
- Drives the product — navigates, filters, and pivots through your Phoenix data the same way you would.
- Investigates failures — walks failing traces and proposes root causes instead of leaving you to grep through spans.
- Iterates on prompts — reads, edits, and tests playground prompts, with every change shown as a diff you approve.
- Reasons over your data — a sandboxed runtime lets PXI query your Phoenix instance to answer questions evals and dashboards cannot.
- Knows the product — Phoenix’s own documentation is wired in as a first-class source, so answers are grounded rather than guessed.
Skills
A skill is a reusable, multi-step procedure for one Phoenix workflow — what to look at, in what order, and what the output should be. PXI loads the matching skill on demand rather than improvising each investigation from scratch. The library is under active development and grows each release. Track progress on the Phoenix roadmap milestone.Trace debugging
Walk failing traces against a failure-mode checklist and return prioritized
root-cause hypotheses.
Span coding
Open-code spans with working notes, then group them into failure categories
and promote the stable ones into structured annotations — a way to build a
taxonomy before one exists.
Span annotation
Annotate spans and apply labels in bulk across a trace or project, and draft
the annotation config itself — categorical, continuous, or freeform. Config
writes are shown for approval before they replace anything.
Prompt playground
Co-author and optimize prompts, with every edit shown as a diff to approve.
Evaluator authoring
Draft and refine LLM-as-a-judge and code evaluators against your data.
Datasets
Turn the failures you find into curated datasets, splits, and labels for
experiments and evals.
Experiments
Run and read experiments over a dataset to compare prompt or model changes.
GraphQL queries
Query the Phoenix API with GraphQL for custom analysis, or get working queries
for your own scripts — built on schema patterns that skip introspection.
You stay in control
Agent assistance is opt-in. PXI can be turned off completely, runs under an explicit permission model, and only reaches the internet when you let it.
Settings → Assistant — system, personal, and web-access controls in one place.
- Turn it fully off. Disable PXI per deployment
(
PHOENIX_DISABLE_AGENT_ASSISTANT=true), per instance (Settings → Assistant → System settings), or per browser (Settings → Assistant → Personal settings → Use assistant). - State-changing actions are gated by approval. Saving a prompt, creating or
replacing an annotation config, writing dataset examples, submitting an
evaluator — each is proposed as a reviewable diff or card and applied only when
you accept, under an edit-approval mode you pick from the chat input (or cycle
with
Ctrl+T). Read-only actions run freely. There is one deliberate exception: while open-coding spans, PXI writes its working notes straight to the span (a single note taggedpxi, updated in place) without a prompt, so it can iterate without asking on every observation — see the span-coding skill under Skills. - Add web grounding when you want it. Toggle web access with the globe button in the chat input to let PXI consult the live internet for additional grounding. The toggle is per session and only appears when an administrator allows it; leave it off to keep the session entirely inside your Phoenix instance.
- Choose where it lives. Keep the assistant pinned to the top navigation bar or switch it to a draggable floating button under Settings → Assistant → Personal settings → Floating assistant button.
Setup
Configure credentials for at least one provider via environment variables or Phoenix secrets (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, AWS
Bedrock credentials, or a custom provider under Settings → Models). The
consent gate shown the first time you open PXI enables the chat surface for that
browser; it does not override system settings.
PXI relies heavily on tool calling — almost every action it takes is a tool
call. Models that are weak at tool use produce broken sessions even if they
handle free-form chat well. Pick one of these validated models unless you have a
specific reason not to:
- Anthropic —
claude-opus-5,claude-opus-4-8,claude-opus-4-6,claude-sonnet-4-6 - OpenAI —
gpt-5.5,gpt-5.4,gpt-5.4-mini - Google —
gemini-3.7-flash,gemini-3.1-pro-preview
How it works
PXI is split between the Phoenix server, which owns everything the model sees (tool definitions, system prompt, skills, capability guidance), and the browser, which executes tool calls that touch the page. Capabilities are gated by context — PXI only advertises a tool when the required Phoenix UI context is present, so it does not offer an action that cannot succeed on your current page.Architecture
Architecture
The server hosts the agent and everything the model sees — its tools,
skills, an MCP client, and a sandboxed bash shell. Bash runs in an
in-process virtual shell on the server with networking disabled; inside it the
phoenix-gql builtin queries the Phoenix GraphQL API directly — the same
schema a logged-in user hits. The browser owns the chat UI and runs the tool
calls that touch the page you are on: applying filters, navigating, and driving
the playground. The server calls your LLM provider with your API key and, when
external resources are allowed, reaches the Mintlify-hosted Phoenix docs MCP.
Runs entirely on your Phoenix
Runs entirely on your Phoenix
PXI runs inside the Phoenix process you are already running:
- Tool calls execute against your Phoenix server and your data — no separate Arize service is involved.
- The LLM is your model provider, called with your API key. Arize is not in the request path.
- Documentation lookups go to the Mintlify-hosted Phoenix docs MCP server when external resources are allowed — the same public docs you can read in a browser, serving docs only.
- Remote trace export happens only if every gate is enabled: a remote collector is configured, an administrator allows export in system settings, and the user enables it in personal settings.
Privacy, safety & configuration
Trace recording & observability
Trace recording & observability
PXI can capture conversations as Phoenix traces, controlled by both system
settings and per-browser preferences. From Settings → Assistant,
administrators can turn assistant access on or off for everyone, allow users to
save session traces locally, and allow export to a configured remote collector;
each user can show or hide the assistant and opt into local or remote trace
recording when allowed.By default the system settings allow neither local persistence nor remote
export. Local traces are written to the
assistant_agent project (override
with PHOENIX_AGENTS_ASSISTANT_PROJECT_NAME). When recording is enabled, tool
inputs and outputs are recorded on the corresponding spans, so you can audit
what PXI did and evaluate it like any other agent in Phoenix.Attach your identity (opt-in). Session traces are anonymous by default. If
you want to associate a session with the signed-in user — for example to
attribute recorded sessions to a specific person — enable Attach your email
to session traces under Settings → Assistant → Personal settings. From
the terminal client the same opt-in is the --attach-user-id flag. It is off
unless you turn it on.Session identity in traces. Recorded sessions carry an OpenTelemetry
session.id so a conversation’s spans group together in the assistant_agent
project. As of Phoenix 20 that id has the shape
{project}:{session-global-id}:{fingerprint}, where the fingerprint is derived
from the session’s creation time. The format changed in this release, so any
saved view, filter, or dashboard that matched the previous session.id shape
needs to be rebuilt against the new one.Safety limits
Safety limits
- Verify before you act. PXI can apply filters, edit prompts, and run bash. Review proposed changes — especially prompt edits — before accepting.
- Bash runs on the server. PXI’s bash tool executes in an in-process
virtual shell on the Phoenix server, with networking disabled and scoped to a
scratch workspace — it does not reach a host machine, a container, or the
internet. Set
PHOENIX_AGENTS_DISABLE_BASH=trueto remove it (see the configuration reference below). - Don’t point it at sensitive production data without controls. PXI sees whatever the signed-in user can see.
- Treat outputs as suggestions. PXI hallucinates, especially on long traces or unfamiliar frameworks.
Configuration reference
Configuration reference
When
phoenix serve starts, it prints the assistant’s effective configuration
as part of the boot banner — whether the agent is enabled, the trace project,
local and remote trace settings, web access, and whether server-side bash is
on — so you can confirm what a deployment resolved to. A malformed
PHOENIX_AGENTS_* value fails the server at startup, before migrations run,
rather than surfacing later at request time.