← Back to Ingenious AI home · All 36 platforms
Ingenious AI · Agent Control
AI agent security · Authorization · Evidence
Control spending, restrict sensitive actions, require human approval and maintain tamper-evident records of your agents' reported activity. Two products do the work: Agent Auth Gateway decides before an action runs, and Agent Flight Recorder preserves the record of what was reported.
Free interactive demonstration — no signup, no API keys, no changes to your systems. By Ingenious AI · Published 9 November 2026
The short answer: an agent you cannot stop, cap or audit is a liability with an API key. Agent Control closes that loop. The gateway checks every routed tool call against spend caps, allowlists, denylists and risk tiers before it runs — allow, deny, or pause for a human. The recorder chains every submitted action into a tamper-evident ledger you can verify after the fact. Around those two sits a wider family of five more platforms — spend mandates, approval records, MCP change gating, log forensics and an agent firewall — mapped below. Every platform sells separately, and every free tier does the complete core job on your real input.
One incorrect tool call can create a real-world consequence. The playground below runs the real policy engine — the same fail-closed code the production proxy uses, evaluated server-side. Paste an action log or pick a preset; nothing is stored.
Refund Assistant
Issue refunds
$100 per refund
$500 refund
DENY — the request exceeds the configured authorization limit, so the call never runs.
Illustrative policy inputs and results — not a claim about an actual customer incident.
An agent loops against a paid API, retrying its way toward a five-figure bill. Spend caps are fail-closed: the call that would breach the cap is denied before it runs, and the reason names the cap.
paid_model_call 24.00 paid_model_call 24.00 paid_model_call 24.00
DENY at the cap
Try it in the playground →An agent attempts a prohibited destructive action. The denylist is absolute — a listed action is denied no matter who or what asks, and on Pro, destructive gates hold irreversible calls for a human or deny them outright.
delete_records 0.00
DENY — denylist
Try it in the playground →An agent calls an approved tool, inside its limits, the way it does a thousand times a day. Policy passes — caps, lists and risk tier all clear — and the action proceeds, logged.
send_email 0.02
ALLOW — within policy
Try it in the playground →An agent keeps trying to act after being disabled. The kill switch is evaluated first, ahead of every other rule: while it is engaged, every later check returns deny, logged, until a human re-enables the agent.
any action, after the switch
KILL SWITCH ENGAGED
Try it in the playground →One honest detail about the demo: the playground evaluates with Pro capabilities, so you can see REQUIRE_APPROVAL verdicts. On the Free plan, unknown and high-risk actions are denied outright instead of held in an approval queue — the same engine behaving honestly at each plan level. Open the playground and read its own FAQ; it says exactly this.
You do not migrate your agent. You wrap its protected tool calls, route them through the gateway, and submit the events you want evidenced to the recorder.
Does the agent have permission to do this? Spending and action policies, allow and deny decisions, human approval, emergency kill switch.
Allowed actions proceed through your own execution code. Denied actions never run. Held actions wait for a human.
What did the instrumented agent report doing? Authorization decisions and action results, hash-chained, verifiable, ready for incident investigation.
POST /v1/actionsWhere the connection stands, plainly. The gateway and the recorder are two separate products today, with separate accounts, and there is no native one-click integration between them yet. The recorder keeps the events your instrumented agent submits — including authorization decisions, if your integration sends them. Automatic forwarding of gateway decisions and execution outcomes into the evidence chain, tied together by one correlation ID, is integration work in progress here, not a feature we will pretend exists. Until it ships, the documented APIs below are the connection — and the assessment wires it into one workflow for you.
The integration is three lines: wrap the tool call you already make, ask the gateway first, obey the decision. From the gateway's own documentation:
Python
gate = requests.post(
"https://authgateway.getingeniousai.com/v1/proxy",
headers={"Authorization": "Bearer YOUR_API_KEY"},
json={"agent_id": 7, "action": "send_email",
"estimated_cost": 0.02}).json()
if gate["decision"] == "allow": send_email(draft)
elif gate["decision"] == "deny": log(gate["reason"])
else: wait_for_human(gate["approval_id"])JavaScript
const gate = await fetch(
"https://authgateway.getingeniousai.com/v1/proxy", {
method: "POST",
headers: { "Authorization": "Bearer YOUR_API_KEY",
"Content-Type": "application/json" },
body: JSON.stringify({ agent_id: 7,
action: "send_email", estimated_cost: 0.02 }),
}).then(r => r.json());
if (gate.decision === "allow") await sendEmail(draft);
else if (gate.decision === "deny") log(gate.reason);
else await waitForHuman(gate.approval_id);Keep development and production keys separate, and never expose an API key in frontend JavaScript, a public repository, or client-visible HTML — the gateway call belongs on your server. Full reference, webhooks and the approval API: the gateway docs · the recorder docs.
These boundaries are printed here, before signup, because control tools earn trust by stating them:
The two products are priced and licensed separately. These are their published tiers, with the limits on the cards where they belong:
1 agent — daily + monthly spend caps, action denylist, kill switch, full evaluation log. No credit card.
10 agents — human approval queue, signed email approval links, webhook notifications, anomaly alerts, burn forecast + auto-throttle, policy simulator.
Unlimited agents — 5 team seats, CSV audit export, full v1 API, anomaly step-up lockdown.
1 agent · 1,000 actions/month — hash-chained ledger, one-click chain verification, embeddable integrity badge. No credit card.
10 agents · 100,000 actions/month — anomaly alerts, incident replay, Tamper-Seal certificates, CSV export + audit report.
Unlimited agents · unlimited actions — 5 team seats, everything in Pro.
How paid activation works today — plainly. There is no instant card checkout yet, and we will not dress the manual path up as one. You order on WhatsApp or by email, we confirm payment — PayPal is the current paid route — and your license key is emailed to you; you redeem it on the product's billing page and the plan unlocks. Crypto payments are coming soon — not live yet. Keys are per-term and nothing auto-renews. A Gateway key does not unlock the Recorder, or the reverse, and there is no combined Agent Control subscription today: a bundle gets priced only when our integration and usage numbers say what it should cost, not before.
Step one is free
The Agent Spend Guard audit is a read-only assessment of your agent spend paths: where your agents can spend, whether a hard ceiling can refuse a call, whether a kill switch exists, whether each agent has a named owner, and whether your records would survive an incident. You get a signed evidence report of the gaps. The audit changes nothing in your stack and stops no spend by itself — the gateway guard is the fix, and it is step two.
The wider family
Seven platforms, one control chain — decide before it acts, prove what it did after.
The gateway and the recorder are the two most teams start with. The other five close the rest of the chain: the spend mandate behind the cap, the approval record behind the decision, a watch on the tools the agent trusts, forensics for after an incident, and a firewall for what the world sends in.
Control tools earn trust by stating what they do not do. Ours, in the open:
Control is in-path. The gateway, the mandate check and the firewall govern the calls routed through them — each is a small, deliberate integration in your agent's loop. An agent that never asks is never checked; wiring the checkpoint in is part of setting the system up, and a kill switch stops future checked actions — it cannot undo actions that already completed.
A hash chain proves integrity, not completeness. It proves a recorded action was not altered after it was written. It cannot prove an action happened that was never recorded, and a receipt is evidence supporting a dispute — not legal protection.
Detection is pattern-based. Injection scoring and log forensics catch known patterns, including encoded and multilingual ones, and both say so on their own pages. No detector catches everything; the policy layer exists for the day one gets through.
If you would rather have the chain fitted to a workflow you already run in production, that is a service, not another subscription: the Production AI Agent Safety Assessment & Retrofit. We map one production agent workflow — its connected tools, permissions and spending paths — identify the sensitive and destructive actions, design the authorization and approval rules around them, wire in the receipt and flight-record evidence, and hand you a written implementation and risk report, including an explicit statement of what is and is not actually enforced. Fixed scope, one workflow — $497, agreed before work starts. Talk to us about an assessment.
What is Ingenious AI Agent Control?
The control-and-evidence journey for production AI agents, built on two products: Agent Auth Gateway decides what an agent may do and spend before it acts, and Agent Flight Recorder keeps a tamper-evident record of the activity the agent reports. Around them sits a family of five more platforms — MandateGuard, Agentic Receipts, TrustGate, BoundaryVault and Rogue-Agent Defense — that complete the chain. Each platform is sold separately and each free tier does the complete core job.
Is this one product or seven?
Seven separate platforms, one family. The two at the centre of this page are separate products with their own sites, accounts and free tiers, and each works on its own. This page exists so a buyer with an agent-control problem can see the whole chain in one place instead of discovering the pieces one at a time.
Do the controls actually stop actions, or just report them?
Both, by design, and the difference matters. The authorization pieces act before execution: a call that breaks a cap, an allowlist or a mandate is denied or held for a human before it runs — for the calls routed through the check. The evidence pieces act after: they record, verify and investigate. A report about an action is not prevention, and this family does not pretend otherwise.
Are the Gateway and the Flight Recorder integrated?
Not natively, yet — and we would rather say so here than let you discover it after signup. They are separate products with separate accounts. The recorder keeps the events your instrumented agent submits, including gateway decisions if your integration forwards them through the documented APIs. Automatic decision-and-outcome forwarding under one correlation ID is integration work in progress; the assessment service wires the two into one workflow today.
How do I pay?
Free tiers need no payment at all. Paid plans are activated by license key today: order on WhatsApp or by email, we confirm payment through PayPal — card checkout is not live yet — your key is emailed to you, and you redeem it on the product's billing page. Nothing auto-renews, and each product is licensed separately.
Which piece do I need first?
Start from the failure. An agent that could overspend: Agent Auth Gateway's free Spend Guard audit, then MandateGuard. A workflow about to face a customer or auditor asking who approved what: Agentic Receipts and Agent Flight Recorder. Agents built on MCP servers you do not control: TrustGate. An agent already running that you could not reconstruct after an incident: BoundaryVault and the Flight Recorder.
Who builds these tools?
Ingenious AI is a software studio in Orange Walk Town, Belize that builds security, compliance and monitoring tools for teams that ship AI systems. Support is at [email protected].