@enforgate/policy

The Enforgate policy engine — the pure, dependency-free core that decides whether an AI agent's tool call is allowed, denied, or needs human approval.

Install

npm i @enforgate/policy

Usage

import { evaluate, type Policy } from "@enforgate/policy";

const policies: Policy[] = [
  {
    id: "p1",
    name: "guard email",
    priority: 10,
    enabled: true,
    rules: [
      { tool: "send_email", action: "require_approval" },
      { tool: "db_*", action: "allow" },
    ],
    defaultDecision: "deny",
  },
];

const verdict = evaluate(
  { serverName: "demo", toolName: "send_email", args: { to: "[email protected]" } },
  policies,
);
// → { decision: "require_approval", reason: "...", policyId: "p1", ruleIndex: 0 }

How it works

• Enabled policies are sorted by priority (desc); rules run in array order; first match wins.
• tool is a glob (*, ?) over the tool name; where adds AND-ed conditions on args (regex matches/notMatches or deep equals).
• If no rule matches, the policy's defaultDecision applies; if no policy decides, the engine default-denies.
• Invalid regexes never match (fails closed).

License

MIT