Paw from Oz

Every time you call an AI API, you're spending money. But unless you're checking the dashboard after each session, you have no idea which project or model is driving the bill.

I solved this by wrapping my AI clients with a one-line middleware that logs every call automatically.

The problem with manual logging

AICostTracker lets you run aicost log myapp gpt-4o 1500 800 to record a call. Useful — but nobody does this consistently. You forget, you're in flow, you'll do it later.

The logs end up patchy and the summary is useless.

Auto-logging with track()

In v0.1.2, I added a track() function that wraps your OpenAI or Anthropic client:

npm install @ozperium/aicost-tracker

Enter fullscreen mode Exit fullscreen mode 

import OpenAI from 'openai';
import { track } from '@ozperium/aicost-tracker';

const openai = track(new OpenAI(), { project: 'myapp' });

// From here, every call is logged automatically
const response = await openai.chat.completions.create({
  model: 'gpt-4o',
  messages: [{ role: 'user', content: 'Summarize this doc' }]
});

Enter fullscreen mode Exit fullscreen mode

That's it. No changes to your existing code beyond the wrap.

It works the same way with Anthropic:

import Anthropic from '@anthropic-ai/sdk';
import { track } from '@ozperium/aicost-tracker';

const anthropic = track(new Anthropic(), { project: 'summarizer' });
const response = await anthropic.messages.create({ model: 'claude-3-5-sonnet-20241022', ... });

Enter fullscreen mode Exit fullscreen mode

What gets logged

After a few sessions:

aicost summary

Enter fullscreen mode Exit fullscreen mode 

  AI Cost Tracker — Summary
  ══════════════════════════════════════════════════
  Total cost:       $1.2847
  Input tokens:     124,000
  Output tokens:    67,500

  By Project:
  ──────────────────────────────────────────────────
  myapp                $    0.9102  31 calls
  summarizer           $    0.3745  12 calls

  By Model:
  ──────────────────────────────────────────────────
  gpt-4o               $    0.9102  31 calls
  claude-3-5-sonnet    $    0.3745  12 calls

Enter fullscreen mode Exit fullscreen mode

How it works

track() patches client.chat.completions.create (OpenAI) or client.messages.create (Anthropic) to intercept the response, extract usage from the usage field, and call logUsage() before returning. It's a thin synchronous wrapper — zero added latency.

If logging fails for any reason, the error is swallowed silently. Your API call always gets its response.

Per-project breakdown

The { project: 'myapp' } option lets you tag every call with a project name. If you're running multiple agents or tools in the same codebase, give each its own project tag:

const researchAgent = track(new OpenAI(), { project: 'research' });
const summaryAgent = track(new OpenAI(), { project: 'summary' });

Enter fullscreen mode Exit fullscreen mode

Then aicost summary shows the breakdown by project automatically.

GitHub: https://github.com/Ozperium/aicost-tracker

Also: AgentSpec for testing AI agent behavior, quota for monitoring rate limits.