The site with no API

What it is and why it matters

The hours you really lose are almost always on a website, not in the code. A social network's studio where you schedule posts, App Store Connect, the CRM, the back office of your management software. Dashboards where you're already logged in, that have no API, or have one that doesn't cover what you need. So every time, you're the one clicking: upload, type, pick, check.

Claude in Chrome solves this. It's a browser extension that Claude Code uses to drive your Chrome, the real one, with your logins. It opens a new tab, reads the page, clicks, types, uploads files from your computer and checks the result. You watch it move in real time. When it hits a login page or a CAPTCHA, it stops and asks you.

We use it every day to schedule STRIKT's carousels on a social network: it uploads the slides, writes the caption, picks the hashtags from the menu, sets the date and time, then checks in the list that the post is there. The secret isn't the extension. It's the procedure written in a file, which the agent reads again every time. This guide is for anyone with a dashboard that eats their hours: by the end you'll have the extension connected and your procedure ready.

1. What you need and how to connect it

• Google Chrome or Microsoft Edge. Claude Code also detects the extension in other Chromium-based browsers, like Brave, Arc, Vivaldi and Opera. It doesn't work in WSL, the Linux inside Windows.
• The Claude in Chrome extension from the Chrome Web Store, on the latest version, and Claude Code up to date.
• A direct Anthropic plan: Pro, Max, Team or Enterprise. You have to sign in to Claude Code with /login: if you use an API key, or Claude through Amazon Bedrock, Google Cloud or Microsoft Foundry, the Chrome integration stays off. Since July 2026 it's no longer in preview.

1. Install the extension and sign in with the same account, and the same organization, you use in Claude Code. If they differ, Claude tells you the extension isn't connected.
2. In the terminal, run claude --chrome. The first time, a dialog explains how site permissions work: press Enter.
3. Type /chrome. It works if you see “Status: Enabled” and “Extension: Installed”. From here you also manage permissions, reconnect the extension and pick the browser if you have more than one.
4. If the extension isn't found on the first try, close and reopen Chrome: it only reads Claude Code's configuration at startup.

In /chrome there's also “Enabled by default”, so you don't type --chrome every time. I leave it off: it keeps the browser tools loaded all the time, and they use up context.

2. Permissions: what it does alone and where it stops

Before an action in the browser, Claude Code may ask for your permission. The prompt starts with “Claude in Chrome wants to” and offers to allow all actions on that site for the session. Site permissions, instead, live in the extension's settings: that's where you decide which sites it can browse, click and type on. On Team and Enterprise plans, the admin can restrict sites with a list.

• It uploads files from your computer: up to 10 MB in total per upload. It refuses files with credential names, like .env, .pem, .key or anything inside .ssh, and the ones it isn't allowed to read.
• Logins and CAPTCHAs: it stops and asks you to handle them. Don't give it passwords: you sign in yourself, once, in your Chrome.
• In long sessions the connection can go to sleep: /chrome, then “Reconnect extension”.

The Claude desktop app also has a built-in browser, but it uses a clean profile, without your logins: the documentation points to it for sites that don't need you, and to Claude in Chrome for acting as you. For us, in the built-in browser, uploading files on the social network's studio didn't work.

3. Write the procedure once

This is the difference between an agent that sometimes gets it right and one you can count on. The first time, you go through it together and you watch. Then you have it write down in a file what worked: click by click, what to check at the end and every trap you fell into. From the next time on, you just tell it “read the procedure and follow it”.

Ours is called PUBBLICARE.md (Italian for “publish”) and it's split into six parts: prepare the files, upload, texts, sound, settings, afterwards. Here's the same structure made generic, to adapt to your dashboard. Replace the parts between < > and remove the symbols.

# Publishing on <site name> (procedure that works)

Tool: Claude in Chrome, Chrome profile <which one>, already logged in as <which account>.
If you hit a login or a CAPTCHA: stop and call me.

## 1. Prepare
- The files are in <folder>. Order: <how they're numbered>.

## 2. Upload
1. Open <exact address of the upload page>.
2. Upload <all the files, in order>.
3. Check in the preview that the order is right.

## 3. Texts
- Title: <maximum length, what goes in it>.
- Description: <rules>.
- Tags: <how to enter them so they really count>.

## 4. Settings
- <date and time, visibility, options to leave as they are>.
- Before the final button, show me a summary: <what can't be changed afterwards>.

## 5. Final check
- Open <the page with the list> and check that <the row with date and time> is there.

## Known traps
- <the trap>: <how to avoid it>.

The traps are the most valuable part, because they're the ones a fresh agent would repeat. These are ours, taken from the real procedure:

• Hashtags. If you paste them all at once, they stay plain text and don't count. You have to type #word, wait a couple of seconds for the suggestions drop-down and pick it with Enter, one at a time. At the end the agent checks that each one has become an element of its own, not text.
• The hidden upload field. Some sites hide the file field behind a button. The procedure says how to make it visible to the agent, so it uploads all the files in order in one go.
• The scrolling time. The time picker is a scrolling list: you click the hour, then the minutes, then you check with a zoom that the field shows the right time.
• A scheduled post can't be edited. On our social network a scheduled post can only be deleted, not corrected. So the check has to happen before the final button, not after.

4. The prompts to copy

Open Chrome on [dashboard address]. I want to [what: schedule a post, update a record, upload a file]. Do it one step at a time and tell me what you see before every click that can't be undone. If you hit a login or a CAPTCHA, stop and call me. Don't press the final button: wait for my OK.

Write in docs/PUBLISH.md the procedure we just did, click by click: the exact address of every page, what goes in every field, what to check at the end and every trap we fell into, with the right way to avoid it. No passwords, emails or keys in the file.

Read docs/PUBLISH.md and follow it for [today's folder]. Before the final button, show me a summary: files uploaded and in what order, texts, tags, date and time. Afterwards, open the list and check that the result is there.

Today this went wrong: [what]. Rewrite the right spot in docs/PUBLISH.md so it doesn't happen again. Don't add a note at the bottom: fix the rule where it is.

5. The site's rules come first

Before automating a site, read its terms of use on automation. Every platform has its own rules, they change over time, and some limit or forbid automated tools. If a site says no, don't do it. I use it on my own accounts, for things I'd do by hand, one at a time, at a human pace.

And it's still a tool acting in your place. Anthropic itself writes that, even with its safety checks, Claude in Chrome carries risks. Give it only the sites you need, keep it away from the ones with your money, and have it stop before every click that can't be undone until the procedure has run cleanly several times.

Common mistakes

• Pasting all the hashtags in one go. They stay plain text: you have to pick them one by one from the suggestions menu.
• Not writing the procedure. Without a file, the agent starts from scratch every time and gets something new wrong every time.
• Trusting the “done”. Without the final check in the list, a failed save looks the same as a successful one.
• Using the wrong Chrome profile. If you have several profiles, write in the procedure which one to use: the agent works with that profile's logins.
• Automating without reading the site's rules. If the platform doesn't allow it, you're putting the account at risk.

Checklist

• The Claude in Chrome extension is installed, with the same account and organization as Claude Code
• /chrome shows “Status: Enabled” and “Extension: Installed”
• I've read the site's rules on automation
• The procedure is in a file, with the final check and the known traps
• There are no passwords, account emails or keys in the procedure file
• The agent shows me a summary before the final button

Frequently asked questions