Features

Eight things the gateway does. None of them live in the prompt.

Every one runs at the tool boundary, outside the model, where the agent holds no keys. A rule in a prompt is a request the model can drop. The same rule in the gateway is enforced whether the model agrees or not.

HoldCanaryOwner routingTokenizationSnapshot and rollbackBatched sendsFreezeConduct grades
Before it runs

Hold for approval

Can't move money nobody approved.

Every action is scored by what it would cost to undo: a new beneficiary, a large amount, a pattern the ledger has seen before. Above the risk gate the action fails closed and waits for the owner of that corridor.

Silence means no. A hold nobody answers expires, and an expired hold never runs.

In the demo: Pay Meridian Consulting LLC $48,200 is sitting in the approvals inbox right now. Open it →

Test-sized canary

At high stakes, the gate does not take the agent's word for the rails.

A test-sized probe goes first. The ledger confirms it landed, then the real transfer follows. If the probe fails, nothing else moves.

The canary is a gateway setting, so the agent cannot skip it.

In the demo: Redwood Compliance Partners, $31,500. Rejected once for a missing W-9, then verified and paid canary-first. See the ledger →

Owner routing

Approvals land with the owner of the consequence, not the person who launched the agent.

Finance holds route to AP. Communication holds route to marketing. Below your tier, the approve button does not exist. An intern can launch the procurement bot; the $48,200 approval still lands with the AP manager.

If the owner does not answer, the hold escalates up the chain and the clock resets. Leavers reroute automatically, vacations delegate, and the top tier can require two people.

In the demo: Maya Chen (Finance & AP) is the approver on every payment hold. Open the approvals inbox →

Tokenization

The model never sees your customers.

PII becomes vault tokens before the model reads it. The tool returns [email protected]; the model sees ⟨customer#4821⟩. The agent cannot leak what it never saw.

Egress is gated like dollars: 3,200 records headed outside fail closed. Every field that crossed the boundary is in the lineage ledger, hash-chained. Residency routing, so EU data stays on EU brains, is on the roadmap.

In the demo: the data rules page shows what the tool returned and what the model saw, side by side. Open the data rules →
After it runs

Snapshot and rollback

Can't destroy what can't come back.

Every destructive call snapshots first: database deletes, file and config changes. Undo is one click inside the window.

"Undo everything" would be a lie. Each action class gets the strongest recovery verb physics allows, and you know which one before the agent acts: reversible for writes and deletes, delayable for anything that leaves the building, compensable for payments and commitments.

In the demo: two stale-record deletions restored from the gateway snapshot, and a payment refunded from the ledger. See the reversals →

Batched sends

Can't email 2,000 people in one breath.

Mass sends leave in batches, stoppable at #50, not #2,000. Outbound email gets an unsend window before anything leaves the building. High-reach publishing is held in escrow, then released batch by batch.

In the demo: Send 6,200 emails, 20% off annual plans, is waiting for Marketing. Open it →

Freeze

Can't spiral.

Three tripwires. Velocity: CUSUM windows add up flows over 24 hours, so five payments of $9,640 get caught as one $48,200, at payment two. Burst: three holds inside an hour and the engine freezes the agent on its own. Arguing: the same action re-submitted after a human said no freezes it on the spot.

A frozen agent gets nothing through. Whatever it still tries is refused in review, and it stays frozen until a person thaws it.

In the demo: ProcurementBot frozen at Apex payment three by the burst rule. The Growth agent frozen by hand at four times its normal volume, thawed after review. Open the kill switch →

Conduct grades

Benchmarks say what a model can do. The ledger says what it did.

Every incident lands in the Near-Miss Report as a conduct grade per agent, A to D: drift off mandate, escalation after a no, velocity, off-hours activity, egress. Trust comes from the ledger, not from the model's report card.

Methodology, thresholds and the grading rubric are public: the Ripcord Index.

In the demo: a week of traffic from four agents, graded. Open the Near-Miss Report →
Gate = P × $exposure × (1 − recoverability)Expected-loss gating. Severity is discounted by our own ability to undo: a snapshot-protected deletion tolerates five times the risk of an irreversible wire.
Σ flows(24h) > h ⇒ escalateCUSUM velocity windows. Aggregates get scored, not just single actions. The structuring attack dies at payment two.
Open the demo workspace → Start free Where the score comes from
← Back to tryripcord.com