@sentinelx/sdk

Pre-execution enforcement for AI agents and automated systems.

SentinelX evaluates action admissibility at the commit boundary, before any irreversible action executes. Every decision returns a cryptographically sealed receipt.

Install

npm install @sentinelx/sdk

Quick Start

import { SentinelX, AdmissibilityError } from "@sentinelx/sdk";

const sx = new SentinelX({ apiKey: process.env.SENTINELX_API_KEY });

// ADMISSIBLE — human oversight present, action may proceed
const receipt = await sx.enforce("ai.agent.action.execute", {
  agent_id:                 "agent_123",
  action_type:              "filesystem.delete",
  human_in_loop:            true,
  human_in_loop_required:   true,
  action_logged:            true,
  harmful_content_detected: false,
  action_within_scope:      true,
});

console.log(receipt.verdict);      // "ADMISSIBLE"
console.log(receipt.receipt_hash); // sha256 sealed receipt

// INADMISSIBLE — human oversight missing, AdmissibilityError thrown
try {
  await sx.enforce("ai.agent.action.execute", {
    agent_id:                 "agent_123",
    action_type:              "filesystem.delete",
    human_in_loop:            false,
    human_in_loop_required:   true,
    action_logged:            true,
    harmful_content_detected: false,
    action_within_scope:      true,
  });
} catch (err) {
  if (err instanceof AdmissibilityError) {
    console.log(err.constraint);     // "AI_HUMAN_IN_LOOP_ENFORCED"
    console.log(err.violation_code); // "INV-026"
    console.log(err.summary);        // "Action blocked: human oversight required at action commitment"
    console.log(err.receipt_hash);   // sha256 sealed receipt — action never executed
  }
}

How It Works

SentinelX sits at the commit boundary between your agent and execution. Before any irreversible action fires, the enforcement engine evaluates it against invariant constraints and returns a deterministic verdict with a provenance receipt.

• ADMISSIBLE → receipt returned, action may proceed
• INADMISSIBLE → AdmissibilityError thrown, nothing executes, receipt returned

The enforcement decision is made server-side. It cannot be bypassed client-side.

API

new SentinelX(options)

const sx = new SentinelX({
  apiKey:    "sx_live_...",
  baseUrl:   "https://enforce.sentinelx.ai",
  timeout:   10000,
  sessionId: "session_abc",
});

sx.enforce(action, context, opts?)

Evaluates action admissibility. Returns a receipt on ADMISSIBLE. Throws AdmissibilityError on INADMISSIBLE.

const receipt = await sx.enforce("ai.agent.action.execute", context);

Options:

• opts.observe: true evaluates without throwing on INADMISSIBLE

sx.evaluate(action, context)

Always returns the receipt. Never throws on INADMISSIBLE. Useful for logging pipelines.

const receipt = await sx.evaluate("ai.agent.action.execute", context);
console.log(receipt.verdict);

sx.verifyReceipt(receipt)

Recomputes the receipt hash client-side and confirms it matches the sealed receipt_hash. This verifies internal consistency and helps detect client-side modification.

const valid = await sx.verifyReceipt(receipt);

sx.health()

const { ok, build } = await sx.health();

AdmissibilityError

Thrown when an action is INADMISSIBLE. Contains the full enforcement receipt.

err.verdict
err.constraint
err.violation_code
err.violations
err.summary
err.trace_id
err.request_hash
err.receipt_hash
err.inv_version
err.receipt

SentinelXAgent

Named helpers for common action types.

import { SentinelXAgent } from "@sentinelx/sdk";

const sx = new SentinelXAgent({ apiKey: process.env.SENTINELX_API_KEY });

await sx.enforceAgentAction({ agent_id, action_type, human_in_loop, ... });
await sx.enforceWireTransfer({ amount, mfa_verified, recipient_verified, ... });
await sx.enforceModelDeploy({ human_approval, model_tested, model_hash, ... });
await sx.enforceOTCommand({ device_id, operator_authorized, mfa_verified, ... });
await sx.enforceBreakerOpen({ operator_authorized, n_minus_1_verified, ... });
await sx.enforceSCADASetpoint({ device_id, parameter, operator_authorized, ... });
await sx.enforceAlgoTrade({ symbol, order_quantity, market_halted, ... });

Receipt Shape

ADMISSIBLE receipt (200):

{
  "verdict":         "ADMISSIBLE",
  "summary":         "Action admissible: ai.agent.action.execute passed all invariant constraints",
  "constraint":      null,
  "constraint_pack": "ai.agent.action.execute",
  "violation_code":  null,
  "violations":      [],
  "mode":            "observe",
  "envelope_class":  "demo",
  "trace_id":        "enf_6be116ac-74a",
  "request_hash":    "sha256:a1ab...",
  "receipt_hash":    "sha256:8955...",
  "inv_version":     "INV-2026-04-02",
  "latency_ms":      0
}

INADMISSIBLE receipt (403):

{
  "type":            "https://enforce.sentinelx.ai/problems/admissibility-violation/human-oversight-required",
  "title":           "Human oversight required at action commitment",
  "status":          403,
  "detail":          "human_in_loop: false — human oversight required but not present",
  "verdict":         "INADMISSIBLE",
  "summary":         "Action blocked: human oversight required at action commitment",
  "constraint":      "AI_HUMAN_IN_LOOP_ENFORCED",
  "constraint_pack": "ai.agent.action.execute",
  "violation_code":  "INV-026",
  "violations": [
    {
      "primitive":  "ai.human_in_loop_enforced",
      "code":       "INV-026",
      "constraint": "AI_HUMAN_IN_LOOP_ENFORCED",
      "message":    "human_in_loop: false — human oversight required but not present"
    }
  ],
  "mode":            "observe",
  "envelope_class":  "demo",
  "trace_id":        "enf_0e7bdaaa-d9b",
  "request_hash":    "sha256:581c...",
  "receipt_hash":    "sha256:351d...",
  "inv_version":     "INV-2026-04-02",
  "latency_ms":      0
}

Domain Coverage

SentinelX enforces actions across high-consequence domains where irreversible execution without constraint validation carries the highest operational risk.

| Domain | Example Actions | |--------|----------------| | AI/ML Agents | ai.agent.action.execute, ai.agent.deactivate, ml.model.deploy.production | | Financial | wire.transfer.execute, algo.trade.execute, treasury.disbursement.execute | | OT/SCADA | scada.setpoint.change, ot.command.execute, breaker.open.execute | | Grid/Energy | load.transfer.execute, der.curtailment.execute.batch, bess.thermal_limit.override | | Cyber/RMM | rmm.script.execute, rmm.privilege.escalate, vpn.session.establish.exploit | | Mobility | driver.payout.execute, surge.pricing.apply, account.suspend.execute | | Delivery | order.refund.execute, dasher.tip.adjust, merchant.payout.execute | | Identity | session.privilege.escalate, directory.credential.export | | Ransomware Defense | historian.data.delete.bulk, hypervisor.encrypt.execute |

Coverage expands continuously. See sentinelx.ai for updates.

Framework Alignment

SentinelX invariant packs address many of the same control concerns emerging in NIST CAISI discussions around agentic AI governance, including execution boundary enforcement for autonomous systems.

| Framework | Alignment | |-----------|-----------| | NIST SP 800-53 | Access Control (AC), Audit & Accountability (AU), Identification & Authentication (IA) | | NIST AI RMF | Govern, Map, Measure, Manage — runtime enforcement at the action commitment layer | | NERC CIP | CIP-005, CIP-007, CIP-010 — OT/grid action enforcement and change management | | MITRE ATT&CK | Pre-execution blocking of techniques mapped to invariant primitives | | IEC 62443 | Industrial automation and control system security zones and conduits |

Troubleshooting

Getting 400? Missing, invalid, or structurally rejected request fields. Every enforce call requires action and context.

Getting 401? Invalid or missing API key.

curl -X POST https://enforce.sentinelx.ai/generate-key

Getting 403 INADMISSIBLE? An invariant constraint was violated. Check err.violations.

Getting 429? Rate limit exceeded. Generate your own key for higher limits.

curl -X POST https://enforce.sentinelx.ai/generate-key

Get an API Key

Generate a test key for observe-mode evaluation:

curl -X POST https://enforce.sentinelx.ai/generate-key

A demo key for immediate testing is also available at sentinelx.ai.

Links

• sentinelx.ai

License

Apache-2.0