Runtime controls for autonomous AI

Control autonomous AI
before it spends or acts.

SentryMandate sits in the execution path for autonomous AI — every request is evaluated for budgets, model access, tool permissions, loop controls, approvals, and audit before it reaches a provider or tool.

Get Started Free See Pricing ↓
ENFORCEMENT PIPELINE Request Auth API key Kill Switch MANDATE ENGINE economics models tools loop policy approvals audit 6 policy dimensions Budget 3 limits Billing quota Approved ✗ Denied
The Mandate

A mandate is the runtime policy attached to an API key.

On every request, SentryMandate evaluates six dimensions — spend, model access, tool access, loop policy, approvals, and audit — before the request reaches the provider. If a check fails, the request is blocked before execution.

MANDATE $ Economics Per-spend & daily caps AI Models Allow-lists & tiers {} Tools Perimeter & risk tiers log Audit Every decision logged Approvals Human-in-the-loop ∞ Loop Policy Runaway intervention One key. One mandate. Policy enforced on every request.
Why Now

Autonomous AI is moving into production faster than runtime controls are catching up. Recursive tool calls, runaway token burn, and uncontrolled API spend are operational problems now — not future ones.

$52.6B AI market by 2030 82% enterprises deploying by 2026 80% Fortune 500 deploying AI 46.3% CAGR through 2030 2024 2025 2026 2027 2028 Deployment Scale Controls $5.1B $11.4B $24.4B Agents are scaling faster than control infrastructure. The control gap widens every quarter. 6 dimensions per mandate <3s detection loop guardian 5-step pipeline per request Real-time 20 event types Models decide. Agents act. SentryMandate governs execution.
$47,000

Lost in 11 days from a single recursive agent loop. One tool-call cycle, zero controls, total budget burn.

— Towards AI
Only 2%

of companies have adequate runtime controls for AI agents — while 95% have already experienced at least one AI incident.

— Cybersecurity Insiders
40%+

of autonomous AI projects will be canceled by 2027 due to runaway costs and the absence of runtime spend controls.

— Gartner
The Stack

The control layer missing from the AI stack

Models generate intent. Agents turn that intent into actions. SentryMandate sits in the execution path to enforce budgets, tool access, loop limits, approvals, and audit before those actions run.

Foundation Models GPT-4, Claude, Gemini, Llama
↓
AI Frameworks LangChain, CrewAI, AutoGen, OpenAI SDK
↓
Tool Integration APIs, browsers, code execution, payments
↓
Execution Control SentryMandate — spend enforcement, loop containment, tool control, approvals, kill switch
↓
Production Your agents, your budget, your production systems
Loop Guardian

Loop Guardian — Runtime detection for runaway agents

Loop Guardian watches active runs for retries, tool loops, stalled progress, and abnormal cost velocity. It detects runaway behavior in under 3 seconds and can warn, throttle, pause, stop, or restart automatically.

RUN STREAM Event Stream tool_call error result output SLIDING WINDOW 5 min 500 max 3s cycle Repeat signature hash dedup max_count / total 0.30 Error Cycle error hash grouping consecutive failures 0.30 Cost Velocity $/event normalized $10/event = 1.0 0.20 Progress Deficit 1 - mutations/total activity vs progress 0.20 COMPOSITE weighted average 0.0 — 1.0 Safe < 0.50 Warn 0.50 Throttle 0.70 Pause 0.85 Stop 0.95 Notify WebSocket event Rate Limit cooldown enforced Freeze Run status → paused Kill + Auto-Restart budget gate + backoff auto-restart (max N, exponential backoff) every 3 seconds scores 0.0 — 1.0
Real-Time Control

See it live. Intervene immediately.

Stream 20 event types over WebSocket in real time. See approvals, budget events, key changes, and run state as they happen — then intervene immediately from the dashboard.

BACKEND HANDLERS EVENT BUS CLIENT DELIVERY Spend 6 events Approvals 2 events Keys 3 events Budgets 2 events Kill Switch 2 events Runs 2 events Loop Guardian 3 events broadcast::channel tokio::sync · 1024 buffer 20 EVENT TYPES SpendApproved SpendDenied SpendPendingApproval SpendConfirmed SpendCancelled SpendExpired ApprovalGranted ApprovalDenied KeyPaused KeyResumed KeyDeleted BudgetUpdated BudgetAlert KillSwitchActivated KillSwitchDeactivated RunStarted RunCompleted LoopDetected InterventionApplied RunRestarted WebSocket wss://host/ws?token=jwt JWT + JTI check user_id filter mpsc(64) backpressure 20s heartbeat ping 3 concurrent tasks filter · send · recv React Dashboard TanStack Query invalidation Kill Switch Global freeze — all keys paused, all requests blocked instantly connected · auto-reconnect · exp backoff
How It Works

Three steps to runtime control.

1

Bring your API key

Bring the provider key you already use — OpenAI, Anthropic, Google, Mistral, or another LLM provider.

sk-proj-abc...xyz
2

Create a SentryMandate API key

SentryMandate issues a new API key that proxies requests to your provider key and enforces policy on every call.

ap_key_sentry_...mandate
3

Attach a mandate and start

Define your runtime policy once. From there, every request through your SentryMandate key is checked for spend, tools, loop limits, approvals, and audit.

mandate: 6 dimensions → enforced

Every request checked before execution.

Pricing

Simple, transparent pricing

Start free. Scale as your agents do. Every plan includes the mandate engine, Loop Guardian, and the real-time dashboard.

A single runaway agent can burn through thousands of dollars before anyone notices. Runtime controls pay for themselves the first time they stop a bad loop.

Free
For individual builders validating runtime controls. No credit card required.
$0
  • $100/mo spend cap
  • 10K API calls/mo
  • 5 keys / 3 mandates
  • 2 concurrent runs
  • Basic loop detection
  • Real-time dashboard
Get Started
Most Popular
Plus
For small teams running production pilots with custom loop policies.
$49/mo
  • $5,000/mo spend cap
  • 500K API calls/mo
  • 50 keys / 25 mandates
  • 10 concurrent runs
  • Full loop detection + custom policies
  • Priority support
Start Plus
Pro
For teams with meaningful production volume and automated loop recovery.
$249/mo
  • $50,000/mo spend cap
  • 5M API calls/mo
  • 250 keys / 100 mandates
  • 50 concurrent runs
  • Loop auto-restart
  • Webhook alerts
Scale with Pro
Enterprise
For large deployments with higher limits, custom controls, and commercial flexibility.
$749/mo
  • $500,000/mo spend cap
  • 50M API calls/mo
  • 2,500 keys / 1,000 mandates
  • 500 concurrent runs
  • Custom thresholds + webhook alerts
  • Audit export + custom pricing above $500K
Contact Sales
Get Started

Bring your agents under runtime control.

Bring your provider key. Create a SentryMandate key. Attach a mandate. From that point forward, every request is checked for spend, tools, loop limits, approvals, and audit before execution.

Get Started Free →