I built a local-first CLI to track what I'm actually spending on AI APIs

Every month I got surprised by my OpenAI bill.

Not by a lot — but enough that I’d open the usage dashboard, squint at a chart, try to remember which project burned all those tokens, and give up. The dashboard shows totals. It doesn’t help you understand which project cost the most, which model was doing the work, or whether that refactoring session with GPT-4 was actually cheaper than Claude.

So I built AICostTracker — a local CLI that logs API calls and shows a cost breakdown by project and model.

How it works

Log a call after you make it:

aicost log myapp gpt-4o 1500 800 "generated README"
aicost log agentspec claude-3.5-sonnet 2000 1200 "test run"

Enter fullscreen mode Exit fullscreen mode

Then see what you spent:

aicost summary

Enter fullscreen mode Exit fullscreen mode

Output:

  AI Cost Tracker — Summary
  ══════════════════════════════════════════════════
  Total entries:    2
  Total cost:       $0.0358
  Input tokens:     3,500
  Output tokens:    2,000

  By Project:
  ──────────────────────────────────────────────────
  myapp                $    0.0118  1 calls
  agentspec            $    0.0240  1 calls

  By Model:
  ──────────────────────────────────────────────────
  gpt-4o               $    0.0118  1 calls
  claude-3.5-sonnet    $    0.0240  1 calls

Enter fullscreen mode Exit fullscreen mode

Why local-first?

I didn’t want another SaaS with an account and a dashboard. The data lives in ~/.aicost/usage.jsonl — one JSON line per call, greppable, portable, yours.

Pricing is baked in for OpenAI (gpt-4o, o1, gpt-4-turbo, gpt-3.5), Anthropic (claude-3.5-sonnet, claude-3-opus, haiku), Google (gemini-1.5), and local/Ollama models (cost: $0).

Install

npm install -g @ozperium/aicost-tracker
aicost log myproject gpt-4o 1000 500 "first entry"
aicost summary

Enter fullscreen mode Exit fullscreen mode

What’s next

The obvious next step is automatic logging — a wrapper that intercepts API calls so you don’t have to log manually. For now, manual logging is good enough to surface the patterns that matter.

If you use multiple AI APIs and want to know where your money’s actually going, give it a try.

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

Also in the stack: AgentSpec for testing AI agent behavior, and quota for tracking rate limits before they stop you.