Four walkthroughs of Proofpane doing the thing. Pour a coffee — they're not 15-second loops.
Looking for the 15-second teasers? Watch them on the homepage.
Four commands, four different keys — and the agent doesn't get to pick which.
5 minutes
For anyone who has ever clicked Approve without reading it.
An AI agent runs four ordinary commands. Each one lands in a different
authorization tier, and the tier is decided from the command
itself — before the card is drawn, not by the person being
interrupted. Reading a file settles with a click. Writing host state needs
Touch ID. Changing what ships to production needs a
hardware security key. And when the agent reaches for the policy
file that governs it, the action is rendered on a separate signing
device, read there, and signed there.
What it shows
The tier is classified from the command, deterministically — an operator cannot choose the faster path for something that earned the slower one
A tray click cannot settle tiers 2, 3 or 4, no matter who clicks it — the server refuses it
Touch ID for host state, a hardware key for deployment config — a synced passkey is not enough for what reaches production
The agent reaching for its own policy file triggers a full ceremony: the command is rendered on a device the machine doesn't control, and read there before it is signed
Shot in one take on a phone. The commands really execute; the targets are throwaway files — watch in higher definition
Five AI tools read a secret file — one audit, one tray, the key comes back redacted.
2.5 minutes
For CISO, security, and anyone whose team runs more than one AI tool.
Cursor, Claude Code, Codex, Hermes and Claude Desktop each try to read a
secret file. Every call — across every tool — lands in
one hash-chained audit; a menu-bar tray fires the
Approve / Deny checkpoints live; denied reads are blocked before the
secret leaves; and an approved read runs through governed tools so the
AWS key comes back redacted by DLP before the model ever sees it.
What it shows
Five different AI tools — Cursor, Claude Code, Codex, Hermes, Claude Desktop — every call in one hash-chained audit
A menu-bar tray fires the Approve / Deny HITL checkpoints, live — multiple apps queued at once
Denied reads are blocked — the secret never leaves the machine
An approved read runs through governed tools → the AWS access key returns [redacted] by DLP before the model sees it
Each tool governed by the deepest method it exposes — hook, MCP, or broker
AI off the rails? Catch it in Slack — approve the fix in Slack.
1 minute
For CISO, Ops, RevOps, anyone whose team lives in Slack.
Whenever an AI agent or workflow goes off the rails — drift,
hallucination, compliance breach, security flag, anything —
Proofpane catches it and pushes the alert to a desktop
tray notification AND a Slack message. The right approvers
decide whether the AI is allowed to self-repair. And the people
who notice a problem first — even without standing permission
— can request access from the same Slack message, on
the spot.
What it shows
Detects drift / hallucination / compliance breach / security flag — the full anomaly surface
Pushes the alert to both the desktop tray and Slack — one detection, two notification channels
Approver decides in-place whether the AI is allowed to self-repair
First-noticer without standing permission? Request access from the same Slack message, on the spot
One detection surface, two notification channels, one human-gated decision — evidence by construction
Salesforce, HIPAA, and an open-ended build — all audited
3 minutes
For Sales Ops / RevOps, CRO, CISO worried about CRM PII,
and engineering leads evaluating "is this actually a platform?".
Three capabilities back-to-back: pull a real Salesforce lead via
SOQL, run a multi-skill HIPAA compliance review from natural
language, and watch an open-ended "build me a brick breaker game"
request iterate to a working build. Every step audited, observable,
and reusable.
What it shows
Salesforce SOQL pull via the platform's connector — every query + result audited
Multi-skill HIPAA review from natural-language prompt — skills compose into the workflow without rewriting them
Open-ended "build me a brick breaker game" request → assistant iterates → working build, all steps hash-chained
Platform as CLI + assistant — operator directs work in natural language; the chain is the proof
Reusable — every artifact (SOQL query, compliance verdict, build steps) is a first-class object the next operator can replay