Agent Action Guard

Framework to block harmful AI agent actions before they cause harm — lightweight, real-time, easy-to-use

Install

npm i agent-action-guard

If you prefer pnpm:

pnpm install agent-action-guard

Usage

import {
	actionGuarded,
	ensureActionSafety,
	isActionHarmful,
} from 'agent-action-guard';

const action = {
	type: 'function',
	function: {
		name: 'send_email',
		arguments: {
			to: '[email protected]',
			subject: 'Status update',
			body: 'Hello',
		},
	},
};

const { label, confidence } = await isActionHarmful(action);
if (label) {
	throw new Error(`Blocked: ${label} (${confidence.toFixed(2)})`);
}

// --------- Create a guarded version of the function ---------

const guardedSendEmail = actionGuarded(
	async function sendEmail(params) {
		return `sending to ${params.to}`;
	},
	{ confThreshold: 0.8 },
);

await guardedSendEmail({
	to: '[email protected]',
	subject: 'Status update',
	body: 'Hello',
});

See examples/basic-usage.js for a minimal runnable example.

Required environment

The classifier expects embeddings from the same embedding model used during training.

export EMBED_MODEL_NAME="sentence-transformers/all-MiniLM-L6-v2"
export EMBEDDING_BASE_URL="http://localhost:1234/v1"
export EMBEDDING_API_KEY="your-embedding-key"

EMBEDDING_BASE_URL should point to any OpenAI-compatible embeddings endpoint. EMBEDDING_API_KEY falls back to OPENAI_API_KEY.

Notes

• This folder implements only the Action Guard runtime.
• Benchmark, dataset, and training code remain in the Python side of the repository.
• The ONNX model expects a 384-dimensional embedding vector.
• The model training script is at ../python/training/.