Agent Access Sandbox

An AI agent works for Dana, a made-up customer. It can reach her bank, email, calendar, and CRM, but only through a broker. The broker holds the login tokens and checks every request against Dana's rules. Click a request to send it. Change a rule and send it again. Watch what gets through.

Everything here is fictional and runs in your browser. Nothing is sent anywhere. It's modeled on the broker I run in front of my own agents: reads go through, and changes wait for me. How I build with Claude.

Agent Requests

    Dana's Rules

      Waiting for Dana

      Token vault

      The broker holds these. The agent has seen 0 of them.

      • Bank: read, pay
      • Email: read, send
      • Calendar: read
      • CRM: read, edit

      Audit Log

      0went through
      0asked Dana
      0blocked
      $0moved today
      #RequestResultWhy

      Nothing yet. Click a request.

      Why Build It This Way

      Agents that act for customers raise a hard question for any company that holds money: what may an agent do on its own, and what must wait for a person? This sandbox shows one answer. Keep the tokens out of the agent's hands. Let reads flow. Put every change that costs money, or can't be undone, in front of the owner. Treat any instruction that arrives inside an email or a web page as suspect. Log everything.

      Try this. Turn off "Ignore instructions from emails and web pages," then send the $9,900 wire. The daily total still blocks it. Turn that rule off too, and the wire still lands in front of Dana, because she has never paid that account. No single rule carries the whole load.

      Want to see the real broker? Email chris@cmm.dev and I'll walk you through it.