Guardrails

Guardrails add a safety and validation layer at three points in the agent lifecycle.

Input Guardrails

Run before the agent loop starts. Can block, modify, or pass through the user query.

import type { IInputGuardrail } from "mitm-ai";

const profanityFilter: IInputGuardrail = {
  name: "profanity-filter",
  async run(input) {
    if (input.includes("badword")) {
      return { action: "block", reason: "Profanity detected" };
    }
    return { action: "pass" };
  },
};

// Or modify the input before it reaches the agent:
const sanitizer: IInputGuardrail = {
  name: "sanitizer",
  async run(input) {
    return { action: "modify", modified: input.trim().toLowerCase() };
  },
};

Output Guardrails

Run on the final OUTPUT content before it's returned.

import type { IOutputGuardrail } from "mitm-ai";

const redactSecrets: IOutputGuardrail = {
  name: "redact-secrets",
  async run(output) {
    const redacted = output.replace(/sk-[A-Za-z0-9]+/g, "[REDACTED]");
    return { action: "modify", modified: redacted };
  },
};

Tool Guardrails

Run before every tool executor call. Inspect or block tool usage.

import type { IToolGuardrail } from "mitm-ai";

const noDestructive: IToolGuardrail = {
  name: "no-destructive-commands",
  async run(toolName, input) {
    if (toolName === "execCli" && /rm\s+-rf/i.test(input)) {
      return { action: "block", reason: "Destructive commands not allowed" };
    }
    return { action: "pass" };
  },
};

Registering Guardrails

Agent.builder()
  .addInputGuardrail(profanityFilter)
  .addInputGuardrail(sanitizer)       // multiple guardrails run in order
  .addOutputGuardrail(redactSecrets)
  .addToolGuardrail(noDestructive)
  .build();

GuardrailResult

Every guardrail returns one of three actions:

action	effect
{ action: "pass" }	Allow through unchanged
{ action: "block", reason: string }	Abort run, return blocked message
{ action: "modify", modified: string }	Replace with modified value and continue