Concepts

AI Assistant

Build and operate agents, flows and knowledge bases in plain language, with every change shown to you before it runs.

The AI Assistant is a panel inside the app that works in one project at a time. You describe what you want — "create an agent called Helper with a friendly support prompt", "what flows do I have?", "add this policy to a new knowledge base" — and it reads, builds and runs things for you using the same capabilities that external AI clients get over MCP.

Where it lives

  • Open it from the AI Assistant tab on the right edge of the screen (the chat button in the header on small screens). Whether the panel is open is remembered in your browser, so it is still open after a reload.
  • Inside a project it chats about that project. Anywhere else (for example the dashboard) the panel lists your projects, most recently active first — the latest of a project's last edit and its last run. Selecting one opens that project, and the panel switches to its assistant.
  • Platform administrators do not see the assistant: it acts on behalf of an organization member.

The conversation

  • Answers stream in as they are written and are rendered as Markdown. Press Enter to send and Shift+Enter for a new line.
  • The conversation belongs to the project and your browser tab. It survives a page reload, a new tab starts empty, and the server stores no chat history. New chat clears it.
  • The assistant replies in the language of your latest message, even if the conversation started in another one. See Language below.
  • Before a task it loads a short guide (skill) for that kind of work, such as building flows or setting up agents, and it copies the names, prompts and descriptions you type exactly as written. It always reasons at high effort.
  • It can read as much as you are allowed to see, but does at most twelve steps per message. If it cannot finish in that many it stops and tells you what it ran and whether anything was changed; send another message (for example "continue") and it carries on from there. Long lists are cut to the items that fit and the assistant is told how many were left out.
  • A conversation can outgrow what the assistant's model can read at once. The assistant then says the conversation is too long (rather than failing silently); start a new chat.

Changes need your approval

Reading — listing and opening agents, flows, knowledge bases, searching the documentation — happens automatically, and so does a flow dry run (it only checks a change and saves nothing). Anything that changes something, and every agent or flow run, is first shown as a card with the exact values it will use. Nothing happens until you press Approve; Reject (or simply typing a new message) cancels it and the assistant carries on. After you decide, the card stays in the conversation marked Approved or Rejected.

You are only asked about changes that can actually run. If the assistant proposes something its arguments do not allow, or something your role does not permit, it is refused straight away and the assistant is told why — you are not asked to approve it.

Auto-approve changes skips the cards for everything except permanent deletions, which always ask. The setting is kept with the conversation, so it survives a reload but is switched off again by New chat.

The assistant can create and update agents, flows and knowledge bases, restore earlier versions, add or replace text documents, register remote MCP servers, and select an LLM for an agent from the providers you have already configured. It cannot create providers, delete agents, flows or knowledge bases, and it never asks for or repeats secrets — API keys and credentials are always entered by you in the UI. When you ask it to delete something, it gives you the exact link to the page where you delete it.

Open pages follow its changes

Once a change has run, lists and any open flow or agent refresh. An editor with no unsaved edits shows the new content straight away; if you have unsaved edits the editor keeps them and the usual revision conflict dialog appears when you save. Every change the assistant makes is a new entry in the resource's History, labeled Assistant, so it can always be undone by restoring the previous version.

Language

The assistant replies in the language you write in. Before a reply is shown, the platform confirms that the opening of the reply is in your language with a short, separate model call. If it is not, the reply is discarded and generated again (at most twice); the text is shown only after it passes and then streams normally. If the check cannot decide or is unavailable, the reply is shown as it is. These check calls are counted in your organization's usage like any other assistant call.

Availability and quota

The assistant runs on a model operated by the platform, not on one of your providers. A platform administrator enables it per organization by assigning a model and a monthly token quota; an organization without its own assignment uses the platform's default model and the default quota. Without any usable model the panel tells you the assistant is not enabled yet.

The quota is counted per organization and calendar month (UTC), across all projects and members, and includes the language checks. When it is used up the assistant answers with a quota message until the next month. Platform administrators see calls, tokens and estimated cost per organization and month in Assistant usage.

Permissions and safety

  • The assistant can never do more than you can do yourself in the project: each action is checked against your permissions at that moment, so reducing your role takes effect immediately.
  • Text it reads — tool results and the contents of documents — is treated as data. Instructions hidden inside it are ignored, and anything it then tries to change would still appear on an approval card.
  • Changes are recorded in the project's audit trail as made by you through the assistant.

On this page