Build your own AI agent, safely, in Claude

AI agents have gotten good enough to do real-world tasks for you, book a table, fill a form, compare prices, run a weekly research job. The problem is the popular way to get one is a sketchy open-source tool that runs with full access to your computer, and earlier in 2026 tens of thousands of those got left exposed and leaked people's passwords. You do not need that risk to get the upside.

You can do all of it inside Claude, safely. The key feature is Claude in Chrome: it drives your own browser, does the clicking and typing for you, and pauses before anything sensitive like confirming or paying. Nothing happens behind your back. Below is the safe setup, the starter prompt, and three more tasks to grow into once you trust it.

The safe setup

1. 💳 Be on a paid Claude plan and install Claude in Chrome. Claude in Chrome is available on all paid plans (Pro and Max). Install the extension so Claude can act inside your own browser, on your own logins.
2. 🔑 Stay logged in to the sites you'll use. Sign in to OpenTable, your airline, your shop, whatever the task needs, in that same Chrome. Claude uses your existing sessions, so you never hand it a password.
3. ✋ Ask for the task and watch it pause. Claude navigates, fills the form, and stops before anything irreversible like confirming a booking or making a purchase, then waits for you to approve. That pause is the whole safety model.
4. 📱 Run it from your phone with Cowork Dispatch. Open Cowork in the Claude desktop app, leave your computer awake, and send tasks from the Claude mobile app while you're out. Your computer does the work at home and reports back.

The mega prompt

With Claude in Chrome installed and you logged in to OpenTable, paste this into Claude.

Using Claude in Chrome, go to OpenTable and book me a table for 2 this Friday at 7pm at [restaurant name]. If that exact time isn't available, show me the closest options and let me pick. Fill in my details, but stop and show me the full booking before you confirm anything. Don't complete the reservation until I say go.

3 bonus prompts to run next

Once the dinner booking feels easy, these stretch into real time-savers. Same setup, just stay logged in to the right sites.

How to get the most out of it

1. 🐣 Start tiny this weekend. Book one dinner. Watching Claude fill the form and pause before confirming is what makes you trust it, and once you trust it, you stop doing your own busywork fast.
2. 🔍 Be specific so it doesn't guess. "Book a table" leaves too much open. Give it the size, day, time, and place, and tell it what to do if your first choice isn't available, so it asks you instead of picking for you.
3. 🛑 Keep the pause on for anything money or personal. The stop-before-sensitive-actions behaviour is your seatbelt. Don't tell it to "just confirm everything," especially on purchases, payments, or anything you can't undo. Review, then approve.
4. 🪜 Brand new? Use it for read-and-fill tasks first (looking things up, drafting a form) before letting it transact. Power user? Chain steps in one task (research, then fill, then summarise) and use Cowork Dispatch to kick off the long ones from your phone while your computer works at home.

Why Claude, not OpenClaw

OpenClaw is a real tool, but it runs with full access to your computer, and in 2026 a large number of instances were left exposed and leaked users' credentials. Claude in Chrome is first-party, works inside your own logged-in browser, and pauses before risky actions by default. For everyday tasks, you get the agent upside without handing your whole machine to something unvetted. That trade is the entire point.