KAGI VAULT

Governed Agent Execution For Accelerator-Grade Evaluation

The product argument is simple: agents should ask Kagi to act, not carry raw authority themselves.

Judge Lens

The strongest demo sequence is Unsafe baseline, Protected rerun, then split-screen compare and live integration.

This version persists state across restart with Vercel-safe external storage, supports named native adapters, and has a dedicated side-by-side compare view with judge export.

Live adapters armed: 0/3

Secret

Missing

Mode

Protected

Scenario

Standard Demo

Target

GitHub Actions, Datadog Logs, Slack Webhook

Runtime

Idle

Trust

100

Core Positioning

Kagi Vault is an execution gateway, not a secret form.

The accelerator-friendly framing is execution control. Secrets stay sealed on the server, agents hit one policy-enforced endpoint, logs are visible, trust changes are measurable, and authority can be revoked mid-stream.

What Changed

State survives server restart through Redis on Vercel, with local file fallback in dev.

Named adapters now target GitHub Actions, Datadog Logs, and Slack.

Unsafe and Protected can now be shown in a dedicated compare view.

A judge report exports current readiness, policy posture, and audit evidence.

Security Boundary

Secret stays inside Kagi.

The agent never receives the raw credential. Even native adapters and webhook forwarding stay brokered server-side.

Control Evidence

Every action is visible.

Judges can see exactly which requests executed, which were blocked, and when trust forced revoke.

Integration Story

Replace direct calls with one endpoint.

This is now credible because the proxy can call named systems directly, not just a generic webhook.