Let your agent go into your portals with your session and bring back the data

The agent uses your own browser, already signed in to your bank, your supplier portal or your sales dashboard, in a separate window, and hands control back to you when there is a captcha or a confirmation. You keep working.

Level
Low code
Requires a simple automation or an API call.
Time
One hour to install it and do the first real task
Published

What it costs

The software is free. You need a compatible agent (Claude Code, Cursor or another that can run commands), with its own subscription or usage cost.

guided mode // I ask, you answer

ready to start

Do it with me, step by step

I will ask you 4 questions about your business, use your answers to get the prompt ready to copy, and then walk you through the 5 steps, one at a time. Like a Thermomix: you just answer and move on.

Also from Claude or ChatGPT with the Master Cookbook MCP: say "start the recipe dejar-que-el-agente-use-tu-navegador-ya-logueado".

Who it is for

Administrators and owners who every week log in to five different portals (bank, suppliers, invoicing, sales platforms) to pull the same report or fill in the same form.

The problem

AI agents can reason but cannot get into your portals: they do not have your session, captchas stop them and sharing passwords with them is a terrible idea. So you keep doing the repetitive work yourself.

What changes

You ask the agent "go into the sales portal and pull this week's report" and it does it in a separate window with your session, without you having to stop working; when a captcha or a confirmation appears, it asks you to take control and then continues. According to the README, BrowserSkill connects agents such as Claude Code or Cursor to the already signed-in browser and comes with the pause for human intervention built in.

Ingredients

  • An agent that runs commands: Claude Code, Cursor, Codex or another compatible one
  • The browser where you already have your sessions signed in
  • The repository's CLI and extension (bsk)
  • A repetitive read-only task to start with: pulling a report, checking a status

Step by step

  1. 1

    Choose a read-only task

    The first task never writes or pays: "go into portal X and bring me the balance" or "download this week's sales report". That way you learn the flow without risk.

  2. 2

    Install the CLI and the extension

    Follow the README for your system. The extension installs in the browser you use every day; the CLI is what the agent runs.

  3. 3

    Open a session and ask for the first task

    From the agent, ask for the task in plain language. You will see a separate window where the agent browses with your session. According to the README, if a captcha or a login appears, it asks you to take control and then continues.

  4. 4

    Verify the result against the portal

    During the first week, compare what it brought you with what you see yourself. When it matches ten times in a row, trust it.

  5. 5

    Write the task down as a routine

    Save the exact request that worked. Next time it is one sentence. If you want, schedule a weekly reminder to trigger it.

The prompt

Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.

# Role
You are my operations assistant at [COMPANY NAME]. You have access to my browser with my sessions signed in, through the connected browser tool. You only read and download; you never submit payment forms or change settings without my confirmation.

# Context
The portal is [PORTAL NAME AND WHAT YOU USE IT FOR, FOR EXAMPLE "the Mercado Libre sales dashboard"]. The data or report I need: [WHAT YOU NEED TO PULL]. Frequency: [HOW OFTEN YOU NEED IT].

# Task
Go into the portal with my session, navigate to the indicated data or report, download or copy it, and give it back to me in an orderly way. If a captcha, a login or any confirmation appears, stop and ask me to take control; when I hand it back, continue.

# Output format
First, the requested data or table. Then, one line with the date and time you pulled it and the exact URL it came from. If something does not match what was expected (for example, the report is empty), say so before delivering it.

# Examples
Request: "Pull this week's sales". Expected output: a table with day · orders · amount, followed by "Extracted on 2026-09-22 at 10:14 from https://…".

What usually goes wrong

  • Starting with a task that writes or pays. The first mistake costs you your trust; start by looking.
  • Not verifying during the first week. A portal changes a menu and the agent can bring back the wrong report with total confidence.
  • Automating sites that forbid it. Read the portal's terms; if they do not allow automation, the convenience is not worth the risk of losing the account.

When not to use this recipe

If the site forbids automation in its terms of use, or if the data belongs to third parties and not to you, do not do it. And never let the agent act unsupervised on portals where it can move money: looking yes, paying no.

Source and real case

Administration · portals and reports

The repository, trending monthly on GitHub at the time of review, connects agents such as Cursor, Claude Code or Codex to the user's already signed-in browser, in a separate window, reusing the real session state and with human intervention built in for captchas, logins and confirmations. The recipe sets the safety rules for using it in a small business.

Trending repository:Tencent/BrowserSkill — Let AI agents use your browser without interrupting your work · GitHub

Last verified:Repository README read on 2026-09-22 (reuses the already signed-in session, separate window, pauses for human intervention on captchas and confirmations); the URL returned 200.

AgentsBrowserPortalsAutomation

Share it

If someone on your team could use this recipe, send it in one click. The suggested text copies itself.

Want to apply it in your business? Message Cristián directly.

Telegram