Govern any AI agent from the first tool call
Before your agent acts, MetaMynd returns a verdict — allow, observe, block or escalate, or containment if the agent has been suspended — checking the agent's mandate, the Standards it's bound to, and your SOPs. Every decision is Ed25519-signed, deterministic, and anchored as tamper-proof evidence.
You don't hand MetaMynd your data or your code — you hand it decisions to be checked. Wrap a tool with the guard and it runs only when the gate allows; the rules live in MetaMynd and a compliance owner can change them live, with no redeploy of your agent.
The flow every action follows
Your agent wants to act → it sends a signed authorize request → the gate checks the mandate + Standards + SOPs → allow (do it) · block (refuse) · escalate (hold for a human) → the decision is anchored as evidence.
What you get
- One-call provisioning → a portable config
- Zero-dependency guard for any Node agent
- allow / observe / block / escalate on every governed action
- Human-in-the-loop approvals (escalate)
- Live rule changes — no redeploy
- Tamper-proof, offline-verifiable audit trail
Three steps to your first decision
- Provision an agent (one API call, or the CLI)
- Point the guard at the returned config
- Wrap your tools with guard.guardTool(...)
See it in code
import { createGuardFromConfig } from "@metamynd/agentsafe-guard";
// Load the config the onboarding call returned (identity + mandate + issuer keys).
const guard = await createGuardFromConfig("./agent.metamynd.json");
// Wrap any tool — it runs ONLY if MetaMynd allows.
const bookFlight = guard.guardTool(
"flight-purchase", // = your mandate scope
rawBookFlight, // your existing handler
(a) => ({ amount: a.amount, currency: "USD", merchant: a.merchant,
context: { tool: "book-flight", riskLevel: a.riskLevel } }),
);
// $400 → allow · $600 → block (SOP_SPEND_CAP) · high-risk → escalate