Research behind paywalls with a computer-use agent

Watch

Chapters

Each chapter starts the class video at that moment.

1. Create a researcher bot28:21
2. The difference: a computer the agent controls30:08
3. The paywall problem30:58
4. Logins go through a separate security bot32:53
5. You approve every credential request34:11
6. Keep credentials in their own bot35:56
7. General-purpose vs. specialist tools37:14
8. Hand deep research to ChatGPT38:43
9. Drive your own ChatGPT, Claude and Gemini accounts39:22
10. One bot per account; let it set model and effort41:05
11. Two ways in: connectors or browser use42:32
12. Some sites need a one-time manual sign-in43:33
13. Search, then let it read (slow, so schedule it)44:47
14. Build your own newspaper45:50
15. “Give me a summary of the first article”46:10
16. Recap: access plus the right model46:23
17. The result: highlights from a paywalled article46:46

From the class

What this module covers

The first agenda block was research. An agent that controls its own computer can read the paywalled sources you already pay for, and credentials stay out of the conversation.

• Create a dedicated researcher bot, give it a focus such as business and markets, and a brief style.
• Each bot can have a computer it controls. That is what lets it open a site you already subscribe to, such as a newspaper, instead of stopping at the paywall.
• He doesn't type the login himself. A separate security bot manages credentials. Live, the researcher asked it for a publication login.
• The request runs bot to bot: the researcher asks the security bot, he enters an approval code, and only then does the agent retrieve the credential from the password manager.
• Keep passwords in their own bot, with its own rules. Never paste a password into the chat.
• His main agent is a general-purpose tool. He hands deep research to ChatGPT, set up as a system the bot calls when research is requested.
• His ChatGPT, Claude and Gemini accounts are set up so the agent can drive them. Each was signed in once during setup. After that the agent uses the account's existing projects. It is slower, because it clicks through the site.
• One bot per account. Tell the agent to set the model and effort level. A laptop is a better host than a phone for this setup.
• Two ways to reach your tools: a connector layer, or browser and computer use.
• Some news sites resist an automated sign-in, so the first sign-in may need to be done by hand.
• Computer use is slow, which is why you schedule the research instead of waiting on it live.
• He builds his own newspaper: a feed of the relevant articles from the sources he subscribes to, then a summary of the first article with the highlights.
• Research comes down to access to the sources you pay for, and using the right model for the job.

Try it

Put it to work

Consumer AI plans may not allow bots to drive their websites. This was a personal setup, not a recommendation for company accounts. Check the plan's terms and ask IT first.

Check each publisher's terms for automated access, and keep what the agent pulls for your own use.

1. Pick one paywalled source you already pay for.
2. Store the site's login in your password manager and have the agent request it through your security bot. Approve with your approval code. Never paste a password into the chat.
3. Ask for the latest on one topic from that source, then ask for a summary of the first article with its link.
4. Decide which model does your deep research, and set it up as its own bot or project.
5. When it works, schedule it. That is the next module.

Log into the virtual machine's browser to my [site] account. Request the credential through my security bot and wait for my approval code. Never ask me for a password in chat.

What's the latest news on [topic] in [publication]?

Give me a summary of the first article.

Set up a system that uses ChatGPT to do deep research when this bot is asked for research.

Check your work

You're done when

• Your research bot can open one source you subscribe to without you typing a password into the chat, and it returns a summary with a link.
• You've decided which model does your deep research.

Tools

Mentioned in this module

• ChatGPT — Deep research, driven through his own account by computer use.
• Composio — The connector-layer alternative to browser use.

Also on the class guide

Written before class

• Pre-class guide: research you can trust
• Pre-class prompt: cited research
• Pre-class prompt: second-model check

Between classes

WhatsApp and YouTube

Join the AI Practicum WhatsApp group

Hive Research Institute on YouTube