AI agents on your website

The Invoicebox chat window on your website also works with AI agents: an agent finds an answer in the help center, writes to support and reads the reply on the visitor's behalf, while the visitor sees every step it takes in the same window.

What AI agents in the browser are

An AI agent carries out a person's request right in the browser: it opens websites, clicks buttons and fills in forms. ChatGPT agent, Claude in Chrome and others work this way. Ask an agent to find out from a store where your order is, and it will come to the chat window on the store's website.

What the Invoicebox chat window offers

The window is clear to agents and screen readers

Each message shows its author and time as text, and fields and buttons have labels. States such as «Operator is typing…» and «Message not sent» are shown as text too. An agent finds what it needs by the labels, just like a blind visitor with a screen reader, and doesn't have to guess what a button does from a picture.

WebMCP tools

WebMCP is a standard that lets a page describe ready-made actions to an agent. In a browser with WebMCP, the chat window gives the agent tools to:

find an article in the site's help center, if the site has a help portal;
create a support ticket;
write a message in a ticket;
view the visitor's tickets and read support replies.

Right after the page loads, the agent sees one chat window tool and uses it to get the others. The window stays hidden until the agent sends something.

The agent works only within the visitor's session and sees the same tickets as the person in the chat window of that browser. Everything the agent sends appears in the window. The agent passes the name and email to support only after the visitor has ticked the consent to personal data processing in the window: an agent does not give that consent.

The operator sees a «Via AI agent» badge on the agent's message. The chat window sets this mark on the visitor's side, so it is a hint, not a verification. Do not change money, bank details or access based on such messages — accept confirmation only from the customer in person.

Where it works

For now, WebMCP is supported by Chrome — in the new-feature trial program (origin trial) or with a flag turned on. The Invoicebox chat window already takes part in the trial, so there is nothing to set up on your website. In a browser without WebMCP, the window works as usual, and an agent uses it the same way a person does.

How to turn it on or off

The tools are on by default. The «AI agent tools (WebMCP)» switch is in your account: «Customer support» → «Sites» → your site → the «Widget» tab. Turn it off and there will be no tools, while the regular chat works as before.

The WebMCP trial on your website

The Invoicebox chat window turns on an experimental Chrome feature on your website — WebMCP: with it, an agent sees the chat window tools right after the page loads. The feature works across the whole page, so your own scripts can use WebMCP too.

You can turn it off in two ways:

with the «AI agent tools (WebMCP)» switch on the «Widget» tab;
with the Permissions-Policy: tools=() header on your website.

The trial is temporary and runs until spring 2027. What happens next is up to Chrome.

What to do as the site owner

Tell agents about the chat window in llms.txt

llms.txt is a text file at the root of a website (https://your-site/llms.txt). AI agents and language models read it to learn what the site has and how to use it. Add a block about the chat window to the file, for example:

## For AI agents

- Support: the «Support» button in the corner of the page opens the Invoicebox chat window. The window's fields and buttons are labeled, and an agent finds them by roles and labels.
- In a browser with WebMCP, the chat window provides tools:
  - `invoicebox_support_load_tools` — available right after the page loads; loads the other chat window tools, the window stays hidden;
  - `invoicebox_help_search` — search help articles;
  - `invoicebox_support_create_ticket` — create a support ticket on the person's behalf;
  - `invoicebox_support_send_message` — write in a ticket;
  - `invoicebox_support_list_tickets` — the person's tickets in this browser;
  - `invoicebox_support_get_messages` — a ticket's conversation with support replies.
- The window passes the name and email to support only with consent to personal data processing. The person ticks the consent in the window.
- [Help center](https://help.example.com/): answers to common questions.

Put the address of your help center in the example. If your site has no help portal, remove the invoicebox_help_search line and the help link. If you don't have an llms.txt file yet, start it with a # Site name heading and a short description in a quote (> …), as the llms.txt format requires.

Don't turn off the tools in your site's headers

The chat window is a separate Invoicebox page inside yours. The widget embed code grants it the right to use tools, so you don't need to add anything to your Content Security Policy (CSP). The permissions the widget needs anyway are enough: https://help.invoicebox.ru in script-src, connect-src and frame-src.

If your site sends a Permissions-Policy header with a list of features, keep tools in it for the chat window:

Permissions-Policy: tools=(self "https://help.invoicebox.ru")

With tools=() or tools=(self), the chat window works, but without tools. If the header doesn't mention tools or there is no such header, you don't need to do anything.

What isn't available yet

An agent can't buy anything through the chat window yet: it won't place or pay for an order in the window. This will come together with sales through AI. The visitor attaches files to messages and closes and rates tickets themselves — these actions aren't available to the agent.

Questions? Write to c-support@invoicebox.ru or call +7 (812) 448-08-01 (9:00–21:00 Moscow time).