Private distribution copy of @agentlaw/core (same source, published under the @mounaji_npm scope). Consumers point at it with AGENTLAW_CORE_MODULE.

@mounaji_npm/agentlaw-core

The deterministic policy engine behind AgentLaw: parse, sign, verify and evaluate policy manifests for AI agent tool calls. No MCP dependency, no LLM in the loop — embeddable in any agent host.

import { loadPolicy, evaluate } from "@mounaji_npm/agentlaw-core";

const policy = loadPolicy({ path: "policy.yaml", trustedPublicKeysPem: [pub] });
const verdict = evaluate(policy.compiled, { tool: "create_payment", args: { amount: 250 } });
// → { decision: "approve", ruleId: "pay-decision", ... } — same input, same verdict, every time

What it provides:

• Manifest schema + loader — YAML/JSON policy manifests (agentlaw/v1alpha1), zod-validated, deny-by-default, first-match-wins.
• Ed25519 signing — canonical JSON hashing, sign/verify with keys that never travel with the manifest.
• Evaluator — typed conditions over tool args, glob matching, rate limits, validity windows and session TTLs. Fail-closed everywhere.

Spec: RFC-0001 — policy manifest. For the CLI and MCP gateway built on this engine, see @agentlaw/gateway.

Apache-2.0.