# Ask about your graph

The optional chat panel helps you inspect the current operational graph. It can
read nodes and signals, but cannot edit the map, execute commands, or invoke
named actions. Team chat requires an editor session.

## Connect a provider

Open the **Agent** tab in the right panel. Choose a provider, enter a model
available to you, and supply its connection settings:

- **OpenAI:** enter an API key. Reticle uses the official Chat Completions
  endpoint; custom OpenAI-compatible endpoints are not supported.
- **Ollama:** enter a local endpoint, normally `http://localhost:11434`.
  Only `localhost` or a literal loopback IP is accepted. Load the model in Ollama
  before asking a question.

For Team, localhost means the **daemon host**, not your browser's computer.
Provider keys and endpoint settings are held for requests and are not saved to
the topology. The panel keeps the entered values while mounted; re-enter them
after reloading.

## Ask a concrete question

Try “Which nodes have failing signals?” or “What does this service depend on?”
The model has read-only graph and node tools. Team also supplies bounded signal
history. Model responses can be wrong; inspect the actual check detail and time
before acting.

Each question starts a new backend exchange. Earlier messages shown in the
panel are not sent as conversation history, so include the node name or context
again when it matters.

## Understand the data path

With OpenAI, your question and graph data retrieved by tools go to OpenAI.
Ollama requests go to the machine running Reticle. Team audit entries record
provider and model, not prompts, answers, keys, or returned graph data.

Desktop chat collects fixed checks with custom commands disabled, even when
the workspace is trusted. Team chat reads the daemon's cached graph, which may
include results of enabled custom checks. Published check output is graph data.

Chat does not replace [named actions](checks-and-actions.md) or provide a shell.
