Exact targets and refusal boundaries
Every Original names the target it supports and checks it before touching anything. An unsupported, ambiguous, or unconfirmable target is refused rather than guessed at.
Trust
Your agent acts within the boundaries you approve
BoxFetch checks the target, pauses before mutation, protects secret inputs, records the outcome, and limits cleanup to resources attributed to the run. Humans control access and can revoke it.
Human approval and approved-plan parity
The plan states the actions, the target identity, and the checkpoint before execution. The kit is designed to execute only the approved plan. Unexpected or unconfirmable states stop the workflow or are reported as uncertain.
Protected secrets and redacted output
Secret values move through protected paths. They are kept out of command arguments, logs, run records, and public responses, and the run record stays redacted.
Outcome checks and honest uncertainty
After the change, the Original checks the resulting state where the kit supports it. A provider timeout or an unconfirmable result is reported as uncertain, never as success.
Run evidence and bounded cleanup
Each run keeps redacted evidence of what it inspected, attempted, changed, created, and verified. Cleanup, where supported, is limited to resources the run can attribute to itself, and anything preexisting, changed afterward, or unattributable stays untouched.
Agent access, OAuth grants, and revocation
A claim link contains a public request ID and is safe to show the intended human. The separate polling secret belongs only to the registering agent and goes only in the claim-status POST body. The MCP client follows discovery, PKCE, and human consent. The human never copies an OAuth access token and can revoke the grant or agent. The controlled beta requires human purchase approval.
Limitations and unsupported targets
No software removes every risk. BoxFetch does not promise that every repository, provider account, or outcome will work. Unsupported targets are refused, and uncertain provider results are reported honestly.
Scope and responsibility
BoxFetch does not remove every operational risk. Teams remain responsible for the targets, credentials, approvals, provider accounts, and application behavior they choose. Unsupported targets are refused, and uncertain outcomes require review.