AI CostGuard

The kill-switch for AI agents.

Prevent runaway AI agent loops, token bombs, and budget explosions before they happen.

npm install @salimassili/ai-costguard

Install & Protect (30 seconds)

import { withCostGuard } from '@salimassili/ai-costguard';

const openai = withCostGuard(client, {
  maxTotalCostPerDay: 5
});

What You Get

When your agent spirals:

[CostGuard] BLOCKED LOOP → 12 recursive cycles detected → estimated save: $18.42

A financial save event. Not a vague error.

Why This Exists

AI agents don't fail gracefully. They fail exponentially.

A stuck agent can burn hundreds of dollars in API credits overnight. Not because it's fast. Because it's relentless.

Your agent framework has no kill-switch. Rate limiters don't detect semantic loops. Cost alerts tell you after the money is gone.

This is the missing layer.

Real-World Protection

1. Recursive Loop

Agent repeats the same reasoning forever.

2. Tool Retry Explosion

Agent retries a failing tool endlessly.

3. Budget Breach

Agent burns tokens rapidly until blocked.

API

withCostGuard(client, {
  maxTotalCostPerDay: 10,    // Hard USD limit
  maxTokensPerRequest: 4000, // Per-call limit
  maxRequestsPerMinute: 30,  // Rate limit
  loopDetection: true        // Catch repeated prompts
})

Example

import { withCostGuard, CostGuardError } from '@salimassili/ai-costguard';

const openai = withCostGuard(
  new OpenAI({ apiKey: process.env.OPENAI_API_KEY }),
  { maxTotalCostPerDay: 5 }
);

try {
  const response = await openai.chat.completions.create({
    model: 'gpt-4',
    messages: [{ role: 'user', content: 'Hello' }],
  });
} catch (err) {
  if (err instanceof CostGuardError) {
    console.log('Agent blocked:', err.message);
    // Handle gracefully
  }
}

Express Middleware

import { costGuardMiddleware } from '@salimassili/ai-costguard';

app.use(costGuardMiddleware({ maxTotalCostPerDay: 10 }));

Features

• Loop detection — kills repeated prompts and recursive cycles
• Daily cost caps — hard USD limits, no overages
• Token limits — block oversized single requests
• RPM limits — catch agents hammering the API

Install Before You Deploy

npm install @salimassili/ai-costguard

MIT | GitHub