A different contract

SSH gives a remote principal an interactive operating-system session. That is appropriate when a trusted human administrator needs a shell. An AI agent usually needs a narrower contract: read this file, apply this patch, run this approved command or call this local MCP tool.

Authorization before execution

ADC evaluates a typed invocation against an Agent-specific grant and device-local limits. The request has an identity, target, bounded lifetime and explicit tool arguments. The result is recorded against the same invocation.

  • Different agents can receive different tools and folders.
  • Writes or execution can require approval.
  • Revocation applies without redistributing SSH keys.
  • Side effects produce durable receipts for retry decisions.

When SSH is still the right tool

ADC does not replace emergency administration, arbitrary interactive debugging or mature SSH infrastructure operated by trusted engineers. It is intended for repeatable Agent actions where explicit scope, approval and audit matter more than unrestricted shell flexibility.