AI assistant

The AI assistant is a chat panel built into Componecat. Ask it about your catalog in plain language (“who owns the payments service?”, “what depends on the auth gateway?”, “summarize the docs for checkout”) and it answers from the catalog itself, not from what a model happens to remember. With your go-ahead, it can also make changes for you.

It runs on a model your organization chooses: you bring your own provider and API key. It is included on every plan.

How it works

The assistant uses the same tools that external agents use through Componecat’s MCP server. Whatever an agent connected from your IDE can look up, the assistant can look up too: searching and browsing the catalog, following relationships and running impact analysis, reading documentation and interfaces, checking scorecards, and reading the wiki. When it uses a tool, the panel shows it inline, so you can see what it looked at to reach an answer.

On an entity’s page or a collection’s page, the assistant knows what you are looking at, so “who owns this?” or “what does this depend on?” needs no further explanation. It is told only which page you are on and fetches the details it needs with its tools.

Set up a model

The assistant appears once your organization has at least one enabled chat model. Until then there is no assistant button in the header.

An organization admin adds one under Organization settings → AI Models:

1. Select Add model.
2. Choose the Chat kind and a provider: OpenAI, Azure OpenAI, Anthropic, AWS Bedrock, AWS Bedrock Mantle (Anthropic), or Ollama.
3. Enter a name, the provider’s model identifier, and your API key or other credentials, plus an endpoint where the provider needs one (for example a self-hosted Ollama server).
4. Use Test connection to check the details, then save.

API keys are stored encrypted and are never shown again after you save them.

You can add several models. One chat model is the organization’s default, and it is pre-selected for every new conversation. A model can be disabled without deleting it, which keeps its credentials for later.

Some installations also offer models configured by the operator. These show with an Instance badge: anyone can pick them, but they can’t be edited or deleted from the settings page.

The same page holds embedding models, which power search by meaning. A chat model and an embedding model are set up separately, and each kind has its own default.

Use the assistant

Open the panel with the assistant button in the header. It docks to the right of the page and stays open as you move around the app.

• Pick a model from the picker under the message box. It lists your organization’s models first, then any instance models.
• Previous chats lists your saved conversations in this organization, newest first. Pick one to continue it, or delete it.
• New chat starts a fresh conversation.

Conversations are saved, so you can close the panel or refresh the page and pick up where you left off. Each person sees only their own conversations, and each organization keeps its own.

Approving changes

The assistant can create, update, and delete things in the catalog, but by default it asks first. Ask before edits is on for every new conversation: when the assistant wants to make a change, it pauses and shows a card with the action and its details. Select Confirm to let it go ahead, or Cancel to stop it. Looking things up never needs approval.

You can turn Ask before edits off for a conversation, and the assistant then makes changes without stopping to ask. Leave it on unless you trust the whole conversation: the assistant reads documentation, wiki pages, and other text written by many people, and an instruction hidden in that text could steer it toward a change you did not intend. Approving each change is what catches that.

Related

• Connect AI agents — give agents in your IDE or chat client the same catalog context over MCP.
• Searching — the query language the assistant uses to find entities.
• Plans and limits — what each plan includes.