---
title: "Use the built-in agent"
summary: "Provider setup, saved threads, search and panel explanations."
description: "Configure a model provider and investigate telemetry in chat."
status: preview
read_when:
  - "You want to investigate telemetry with the built-in agent."
---
## Enable a provider

Select one provider and supply its key before starting Pitot:

```sh
export PITOT_AI_PROVIDER=anthropic
export ANTHROPIC_API_KEY_FILE=/run/secrets/anthropic_api_key
```

For OpenAI, use `PITOT_AI_PROVIDER=openai` and `OPENAI_API_KEY_FILE` instead.
The selected provider requires its key. An empty provider disables chat.
Restart after changing settings.

The default models are `claude-sonnet-5-5` for Anthropic and `gpt-6.1-sol` for
OpenAI. Override with `PITOT_AI_MODEL`. `ANTHROPIC_BASE_URL` and
`OPENAI_BASE_URL` select a gateway or alternate endpoint. See
[AI settings](/docs/reference/configuration/ai) for endpoint and model requirements.

Requests send conversation and query evidence to the configured provider and
can incur provider charges. Set usage budgets with that provider or your gateway.
Every built-in role has `agent:run`, including viewers. This permission gates
agent runs and thread APIs. See [roles](/docs/reference/roles).

## Investigate in a thread

Choose **New chat** in the navigation rail. Include a service, time window and
symptom, for example: "Why did checkout latency rise in the last hour?"
Continue in that thread to keep the conversation context. The agent can query
telemetry and build [dashboards](/docs/guides/dashboards).

Threads belong to the signed-in user. Select a saved chat in the rail to reopen
it. Use **Search** at the bottom of the rail, or press **⌘K** (**Ctrl+K** on
other keyboards), to find chats, dashboards and services. A service result
starts an investigation for that service.

Open a chat's menu to **Rename** it or **Delete** it. Confirm deletion in the
dialog. Renaming accepts a nonempty title of up to 120 characters.

## Explain a panel

Open a dashboard panel's menu and choose **Explain in chat**. Pitot opens chat
with the panel context. Explanation requests are answer-only; the agent can
read evidence and suggest corrections, but cannot save changes during that run.

Review the evidence before acting. Narrow the service or time window if a run
hits its step limit or five-minute time limit.