AI agent

The AI agent builds and edits your project from a plain-language description. It reads and writes each document as a whole — the same tree of nodes you see on the canvas — and every change it makes lands on the canvas as ordinary node edits, so it is visible right away and can be undone like any other edit.

Before it finishes, the agent checks the project the way a build does (compile and type check) and fixes the errors it finds, so a reply usually comes with a project that builds. The steps under a reply show what it did: which documents it read or wrote and when it checked the project.

Opening the agent

Click Agent in the editor toolbar. The chat opens in the right-hand sidebar, next to History (only one of the two is shown at a time). The panel belongs to the whole project, not to a single document: switching tabs does not interrupt a request that is running.

The agent has to be configured once for the whole deployment by an administrator — see AI agent in Administration. Until then the panel shows a “not configured, ask an administrator” notice.

You can also describe a single step in plain words and let the agent build it — see Magic node. To let a coding agent such as Claude Code work on the project from your own terminal instead, see Coding agents (MCP).

Chat sessions

Conversations are kept as sessions per project. You can start a new session, switch between them, rename or delete them, and export a session as a JSON file. A new session is named after the beginning of its first message (up to 50 characters) unless you renamed it before sending. The × button in the panel’s header closes the panel. Under each reply the panel shows how many tokens the turn used, and the individual steps the agent took can be expanded.

What it can change

Edit the document that is open, or any other function or trigger document of the project, and object structures — the documents that describe the shape of your data (types only; they never store anything).
Find the right integration for a task by keywords (for example, a Telegram trigger for a bot command) among the integrations available to the project.
Create new documents, and integration instances (credentials may be left blank for you to fill in later).
Work in a project with no documents at all — describe what you need and it creates them as it goes.

Undo. Every document the agent writes is one undo step — a request that wrote the same document twice (for example, once and then again to fix an error the check found) is two steps. Press Undo in the panel to roll them back one at a time. Your own manual edits made afterwards are separate steps. (If an administrator switched the agent to the node tools, one request is one undo step per document it touched.)

Clarifying questions

When a request is ambiguous in a way that changes the structure of the result — for example, which integration to use — the agent stops and asks instead of guessing. Details it can reasonably fill in itself, it does not ask about.

It asks one question at a time, with two to four suggested options, each with a short description.
Click an option to answer, choose Other… to type your own answer, or click Decide yourself to let the agent pick.
While a question is waiting, the main message box is disabled — answer the question first. Your answer continues the same conversation.
The agent asks at most three questions in a row; after that it decides on its own.
The work done before a question and the work done after your answer are separate undo steps.

If you’d rather not be asked at all, turn on the Don’t ask, just do switch in the chat panel header. The agent then makes its own choices. The switch is remembered per project in your browser.