AI coding agents with
human approval

Autonomous coding agents optimize for throughput: give them a task, get back changed files. That's impressive until an agent confidently rewrites something it shouldn't have — and you find out after the fact, from a diff you never saw.

An approval gate inverts the default. Wanju's built-in agent can propose anything and apply nothing: every built-in edit waits as a side-by-side diff in an approvals inbox, across all your projects, and nothing touches disk until you say so. Applied changes are backed up and revertible; stale proposals are blocked from clobbering files you've since edited.

A real gate needs more than a confirm dialog. Wanju's permission broker is deny-by-default with per-project overrides; commands run through Wanju are restricted to an allowlist (your test, build, and lint scripts — not arbitrary shell); credential-like files are unreadable in every direction; and every dispatch, edit, approval, and denial lands in an append-only audit log with an emergency lock switch.

Verification closes the loop: after you approve, your test suite runs automatically and the result is attached to the run. Approval without verification is just vibes.

The built-in gate works with your own OpenAI, Anthropic, or Gemini key (encrypted with your OS keychain, stored only on your machine). See how the whole loop fits together.