# What the agent can and can't do

The agent does the work in your services and files and hands back a finished result. Below is what it uses for that, and what it doesn't do. What it looks like in your field: [Example tasks](/help/examples).

## What the agent works with

- **Connected services.** Marketplaces, CRM, 1C and MoySklad, banks, e-document exchange, mail, calendars and spreadsheets: 115 services in the [integration catalog](/integrations). The agent reads the data itself and makes changes right in the service. How to connect: [Integrations](/help/integrations).
- **Files.** Reads exports, documents, PDFs, scans and photos, and transcribes call recordings. Delivers the result as a file: Excel, Word, a deck, PDF. More: [Files](/help/files).
- **Calculations.** Calculates in code on a cloud computer rather than in its head: joins tables, reconciles exports, draws charts and diagrams right in the chat.
- **The web.** Searches and reads open pages: laws, tariffs, vendor documentation, competitors' prices.
- **Its own browser.** Opens sites with no integration, fills in forms, works in online accounts. Its browser shows in the **Agent's screen** panel, and you can take over with **Take control**.
- **Your computer.** With the desktop app, the agent works with the files and programs of that computer and in your browser. More: [Apps](/help/apps).
- **Images and video.** Removes backgrounds, resizes for a marketplace, draws product cards and banners, makes short videos. How much your plan allows: [Limits](/help/limits).
- **The agent's email.** Each agent has its own address: it receives emails and replies to them. More: [Agents](/help/agents).
- **Work on a schedule.** Repeats a task on its own and sends the result: [Scheduled tasks](/help/tasks).
- **Messengers.** Replies in Telegram, MAX and other messengers and sends its cards and reports there: [Messengers](/help/messengers).
- **Helpers and other agents.** Splits big work between helpers that run in parallel, and can hand part of a job to another of your agents: [Agents](/help/agents).

## What the agent doesn't do

- **It doesn't touch a service's data without a connection.** Access appears only once you connect the service, and only with the rights you granted when connecting.
- **It doesn't change or send things unchecked.** An action that changes or sends something may come as a card and wait for your answer. You set what it always asks about and what it may do without asking: [Approving actions](/help/trust).
- **It doesn't sign in or pay for you.** The agent doesn't read passwords. When a sign-in, a check or a payment confirmation is needed, a **Sign-in needed**, **A check needs passing** or **Payment needs confirming** card arrives, and that step is yours.
- **It doesn't see your computer without the app.** In the cloud the agent has its own computer. Your files and programs are reachable only while the desktop app is running and connected.
- **It doesn't control equipment.** Connecting machines, sensors and cameras is enabled on request, and the agent only reads their data: it sends no commands or setpoints. More: [Equipment](/help/equipment).
- **It doesn't bring your services and personal memory into a chat with other people.** In a group chat with a colleague and in a messenger group, the agent works without them by default. What can be turned on and who sees what: [Group chat](/help/group-chats).
- **It doesn't work without a plan.** There is no free plan; work runs on a plan or the trial week and uses the weekly limit: [Plans](/help/plans).
- **It doesn't plug in your own model by default.** The agent thinks with the product's models; **Auto** usually fits. Your organization's own models are connected on request, in the OpenAI-compatible and Anthropic Messages formats: [How to reach us](/help/contacts).

## The agent makes mistakes too

Check what matters before passing it on: figures in a report, amounts in a document, the list of changes before an edit in a service.

The agent doesn't grade a scheduled task's result itself: a separate check compares it with the task's conditions. If the check finds open items, the report arrives marked **Not confirmed**; if it couldn't check, marked **Not verified**. A regular chat answer doesn't go through this check. More: [Scheduled tasks](/help/tasks).