Agent Approval Gate

Before an AI agent takes a consequential action, the user confirms it one more time on their phone.

The simplest way to stop a typo in an ad budget from approving ₩50,000,000 of spend: the user taps Reject once on their phone. That's what this feature does.

What makes it different

Compared with a Slack bot, an email notification, or ntfy.sh:

Capability	Plain notification	logi Agent Approval
User confirmation	Visible	The exact content the user saw, SHA-256 sealed
Proof of decision	None	Device secure-key signature (non-repudiable)
Identity	Bot token	The same user identity as logi SSO
MFA-bombing defense	None	6-digit match + cool-down + in-flight suppression
Audit log	Varies by bot	Hash chain + WORM trigger + 7-year retention
Standards	Proprietary	OAuth CIBA (RFC) + RAR (RFC 9396) compatible

Scenario

Codex attempts to change the ad budget
        ↓
calls mcp__logi__request_approval
        ↓
logi → push to the user's phone (APNs/FCM)
        ↓
user: [Reject] or [Approve with Face ID]
        ↓
signed decision → returned to Codex
        ↓
Codex proceeds with the real action, or stops

Every state transition in this flow is recorded as a hash chain in a dedicated audit table.

Next steps

• Quickstart — wire it into Claude Code / Codex / Cursor in five minutes
• Setup guide — how users write rules for their agents
• API reference — REST endpoints and response schemas
• Security — threat model and defense layers in detail