Skip to main content

AI assistant

The AI assistant is a chat sidebar built into the Enterprise Cloud UI. It connects to AI models through OpenRouter and can call tools exposed by the MCP servers in your catalog, so you can interact with your infrastructure through natural language without leaving the browser.

Prerequisites​

The AI assistant requires an OpenRouter API key configured server-side during deployment. The key is never exposed to the browser.

Open the assistant​

Toggle the assistant sidebar with the chat icon in the navigation bar, or press Cmd+B (macOS) / Ctrl+B (Windows/Linux). The sidebar opens on the right side of the page.

Select a model​

The model selector dropdown at the bottom of the sidebar lists models from multiple providers (Anthropic, OpenAI, Google, Meta, and others). The list is fetched from OpenRouter at runtime and filtered to only include models that support tool calling.

Select MCP servers and tools​

The MCP Server Selector next to the model dropdown shows how many servers and tools are currently active (for example, "3 Servers / 12 Tools"). Click it to open the server selection dropdown:

  • Toggle individual servers on or off with the checkbox next to each server name.
  • Click the settings icon on a server to open the tool configuration modal, where you can enable or disable individual tools, search tools by name, or use Enable All / Disable All.

When you send a message, the assistant can call any enabled tool. It decides which tools to call based on your message, executes them through the MCP server, and interprets the results in its response.

Send messages​

Type your message in the input area and press Enter or click the send button. The assistant streams its response in real time. During streaming, the send button becomes a stop button that cancels the current response.

You can attach images and PDFs as context for your message (up to 5 files, 10 MB each). Drag files into the input area or click the attachment button.

Chat history​

Conversations are stored in your browser (IndexedDB) and persist across sessions. The Chat history dropdown in the sidebar header lets you:

  • Load a previous conversation.
  • Delete individual conversations.
  • Clear all conversations with Clear All.

Each conversation records the model and selected MCP servers at the time of creation. The conversation title is automatically set from the first message.

How tool execution works​

When the assistant decides to call a tool:

  1. The request goes to the Cloud UI backend (/api/chat), which authenticates your session.
  2. The backend connects to the selected MCP servers and executes the tool with the parameters chosen by the model.
  3. The tool result is returned to the model, which interprets it and generates a formatted response.
  4. A single request can chain up to 5 tool calls before the model produces its final answer.

All tool execution happens server-side. The browser never connects to MCP servers directly.

Visibility​

An administrator can hide the assistant for all users via the Enterprise Manager. See Feature flags in the Cloud UI configuration for details. The assistant is on by default when no Enterprise Manager is deployed.

Next steps​