Textpot

• 

  textpot_architecture diagram

Inspiration

Comment sections are where the real customer sentiment data lives, yet social media platforms are becoming increasingly harder to scrape. They have been setting up robust paywalls (ex. Reddit APIs) and bot detection mechanisms. The only silver bullet is a real user's browser, which can now conduct automated workflows with Gemini computer use.

The user's browser comes with several perks:

• it's already authenticated on the social platforms the user uses
• it can browse social media platforms without getting blocked by the platforms

What it does

Textpot leverages the user's browser to explore social media, with the backend running on Google Cloud Run. Textpot runs the following loop up to 3 times before returning comment data analysis: (1) extension captures screenshot → (2) POST to Cloud Run → (3) Gemini sees screen, returns action → (4) extension executes action via CDP

How we built it

Extension (user's machine) — Opens the popup windows, attaches CDP, executes the actions (clicks, scrolls, keypresses), captures screenshots. Runs entirely in the user's Chrome.

Cloud Run (backend) — Receives screenshots, calls Gemini, stores the conversation history in memory across turns, returns the next action. Never touches the browser directly.

The split means the browser is already authenticated with the user's real sessions and cookies — Cloud Run just tells it where to click.

Challenges we ran into

The first iteration of the product was not a Chrome extension. It was a web app that tried to browse social media platforms with a headless browser on Google Cloud Run. The result was blocked access due to bot blocking and authentication issues. That's why we pivoted to a Chrome extension.

What we learned

I learned how powerful Gemini computer use is. It doesn't require prompting to navigate to the correct part of a webpage. It's able to correctly view comment sections regardless of what social media platform it interacts with. Though prompting could theoretically boost its performance, I find it super useful that its vision capabilities are truly platform-agnostic.

What's next for Textpot

It'd be nice to launch Textpot on Chrome Web Store someday. Since it partially runs on Google Cloud Run, I'll need to come up with a pricing model.

Built With

• chrome-devtools-protocol
• fastapi
• gemini-2.5-flash
• gemini-computer-use
• google-artifact-registry
• google-cloud-build
• google-cloud-run
• google-gemini
• google-genai-sdk
• google-secret-manager
• javascript
• next.js
• python
• typescript