AI Assistant

The AI Assistant is a widget located in the bottom right corner of the admin panel. In its collapsed state, it appears as a round button; clicking on it opens the assistant panel. While the assistant is working, the sphere pulses, indicating that the agent is still performing a task, even if the panel is closed.

The panel has two tabs (in the English interface — "Documentation" and "MCP Agent"):

Tab	What it does
documentation	answers questions about the OneEntry documentation — does not change anything in the project
MCP agent	performs a task described in plain language: finds the necessary Admin API operation, creates and modifies project data

The tabs have separate histories: questions to the documentation do not enter the conversation with the agent, and vice versa.

Running the agent costs money

Each time the agent is run, it consumes resources from the selected AI account in the panel: the agent queries the model multiple times for a single task — it searches for the operation, parses it, calls it, and checks the result. The cost for each run is displayed under the response, and the total balance is shown on the Settings → AI Agents tab.

Assistant Panel

In the header of the panel, to the right of the tabs, there are three buttons:

• expand / collapse — makes the panel wider or returns it to its normal size;
• clear history — clears the conversation of the open tab. The button is inactive while there is nothing in the conversation or the assistant is still responding;
• close — collapses the panel back into the sphere. The panel can also be closed by clicking anywhere outside of it.

MCP Agent Tab

The agent performs a task described in its own words — for example, "create a page about delivery" or "create products from this table." It works with the Admin API through a set of tools: it searches for the necessary operation, parses its parameters, calls it, and uploads files. Among the tools is a search function for the OneEntry documentation — the agent reads the documentation itself if it needs to understand how the module works.

The agent's permissions are your permissions

The agent acts on behalf of the current administrator and with their permissions: what is unavailable to you is also unavailable to the agent. The agent is prohibited from deleting data, and it can only perform protected operations after your confirmation (see below).

Selecting the Agent and Model

At the top of the tab, in the Agent and Model fields, you select the AI account and model that will perform the task. Accounts are taken from Settings → AI Agents. The list includes only models capable of calling tools: without this, the agent cannot work with the API.

If there are no AI accounts in the project, instead of a selection, the panel shows a warning that the AI agent is not configured and needs to be set up in the settings. Create an account on the Settings → AI Agents tab.

Starting a Task

1. Describe the task in the request field.
2. If necessary, attach a file using the paperclip button — for example, a table with products. Only one file can be attached to the request. The server sets the allowable file types and maximum size; if the file is too large, the panel will not accept it and will suggest splitting it into parts. An image can only be attached if the selected model can read images — otherwise, choose another model or attach a document.
3. Click the start button (in the English interface — "Start"). While the agent is working, the button changes to a cancel button ("Cancel") — it stops the execution.

While the task is being executed, the agent's steps are visible in the feed: which operation it is searching for, which it is calling, which file it is uploading. The step with details can be expanded by clicking. When the agent finishes, its response appears under the steps, and below the response is the cost of the run: the number of tokens and the cost in dollars.

The agent takes into account several of the most recent completed exchanges in the conversation, so the next task can be formulated as a continuation of the previous one.

Confirming Protected Operations

Before a protected operation, the agent pauses and shows a card requesting permission. It includes the method and address of the request, the data that will be sent, and what currently exists at that address. The decision is made based on the request itself, not on the model's summary:

• the allow button (in the English interface — "Allow") — the agent performs the operation;
• the deny button ("Deny") — the operation is not performed.

If the response does not reach the server, the card will inform you of this, and the buttons will become available again.

Important Things to Know

• Closing the panel does not cancel the task. The agent continues to work, and the steps and response will appear when the panel is reopened. However, after closing the panel or switching tabs, it is not possible to stop an already running task from the panel, and any unsent request text and attached file will be lost.
• History is stored only in the page's memory. After reloading the page, the conversation with the agent is lost. Closing the browser tab or reloading while the agent is working interrupts the task.
• If the AI account limit is exhausted or the selected model is not suitable, the task will not start — the panel will show the reason.