Arkangel.AI

A real-time AI assistant that helps parents protect their kids online by using Claude to detect harmful or distracting content, and explain interventions with empathy and real-time context.

Comment

Inspiration

Parental control tools today are either overbearing or ineffective, blocking everything or doing nothing. We wanted to create a digital guardian that doesn't just restrict, but understands. Inspired by Black Mirror’s “Arkangel,” we built a tool that uses Claude’s ethical reasoning to gently guide kids online, while giving parents real visibility, without sacrificing empathy.

What it does

Arkangel utilizes the newest Claude API features including visual image processing and google workspace integration to prevent harmful and distracting content from reaching the user. Arkangel periodically scans the user's screen, identifying inappropriate or off task content, and immediately uses web-ui MCP to redirect the user back on task, and presents a gentle nudge to the use to stay safe.

Features: Real-Time Content Detection: Feeding screenshots ~1 sec to claude api, detecting distractions/unsafe content based on parents’ input

Claude-Generated Ethical Nudges: automatically opening a new tab with a popup explaining Claude's decision-making and nudge to return to task

Context-Aware via MCP + Google Classroom: MCP Integration with Google Classroom API to incorporate additional info about assignments, deadlines, etc

How we built it

Used Python to capture screenshots every second Sent those screenshots to Claude’s API for real-time analysis Used Claude + Browser Use to blur screens, pop up text, and redirect tabs Connected to a Google Classroom MCP server, enabling Claude to reason about active coursework Built a Tailwind frontend for parents to set rules and see flagged sessions Created a Claude-compatible tool manifest to support conversational prompts like "What assignments are due?"

Challenges we ran into

  • Getting the MCP server authenticated with Google Classroom and configuring it for Claude to recognize
  • Managing the asynchronous timing of screenshots and Claude responses in real-time
  • Debugging Browser Use actions when multiple instructions fired at once
  • Making nudges feel empathetic and clear, not annoying or robotic

Accomplishments that we're proud of

  • Attending our first-ever 2 hour hackathon
  • Fast Integration between Claude and real-time browser actions

What we learned

  • How MCP can make Claude so much more powerful
  • Computer Use at the baseline uses screenshots every second + awareness of where the user is moving around the screen

What's next for Arkangel.AI

  • More MCP integrations (Canvas, Google Workspace, Clever)
  • Claude + Vision API to fully replace screenshots with direct scene understanding
  • Streaming input pipeline for continuous, real-time context without polling
  • Customizable Claude alignment profiles (e.g., strict vs. flexible modes)
  • Chrome extension or native desktop agent for deployment in real classrooms

Built With

Share this project:

Updates