llm-tax

Scan your codebase and find out how much money you're wasting feeding bloated API responses to LLMs.

Most APIs return 2-5x more data than your AI agent actually uses. Every extra field is wasted tokens — and wasted money.

llm-tax scans your codebase, finds every API call, hits the endpoints, measures the responses, and tells you exactly how much you're overpaying.

Quick Start

npx llm-tax ./src

That's it. No install, no config, no signup.

What It Does

1. Scans your .ts, .js, .py files for fetch(), axios, requests, httpx calls
2. Extracts which response fields your code actually uses (static analysis)
3. Hits each API endpoint and measures the response size
4. Calculates token waste, dollar cost, and monthly estimates across GPT-4o, Claude, etc.
5. Shows a color-coded report with your worst offenders

Example Output

  ╔═══════════════════════════════════════╗
  ║          💸 llm-tax                     ║
  ╚═══════════════════════════════════════╝

  APIs found:        17 (17 reachable)
  Avg waste:         76%
  Tokens per call:   6610 total, 3474 wasted

  Per-API Breakdown
  ────────────────────────────────────────────────────────

  GET https://api.stripe.com/v1/charges
  ./extensions.test.ts:37
  ████████████████████ 99% waste  │  86 tokens → 1 used, 85 wasted

  GET https://api.openai.com/v1/chat/completions
  ./extensions.test.ts:9
  ████████████████████ 98% waste  │  40 tokens → 1 used, 39 wasted

  GET https://api.github.com/user
  ./_core/sdk.ts:57
  ███████████████████░ 97% waste  │  30 tokens → 1 used, 29 wasted

  Monthly Cost Estimate (1000 calls/day)
  ────────────────────────────────────────────────────────
  Model                Total Cost      Wasted          Savings
  gpt-4o               $495.75         $260.55         53%
  gpt-4o-mini          $29.75          $15.63          53%
  claude-sonnet        $594.90         $312.66         53%
  claude-haiku         $158.64         $83.38          53%

  WORST OFFENDER  GET https://api.stripe.com/v1/charges
  99% waste — 85 tokens thrown away per call

Options

npx llm-tax ./src                    # Scan a directory
npx llm-tax .                        # Scan current directory
npx llm-tax ./src --calls 5000       # Estimate at 5000 calls/day
npx llm-tax ./src --json             # Output as JSON (pipe to file)
npx llm-tax ./src --dry-run          # Scan only, don't hit APIs
npx llm-tax ./src --json > report.json  # Save report

What It Scans

| Pattern | Languages | |---------|-----------| | fetch("https://...") | JS, TS | | axios.get("https://...") | JS, TS | | requests.get("https://...") | Python | | httpx.get("https://...") | Python | | url: "https://..." | Config objects | | "https://api.*/..." | Any API URL |

How Token Waste Is Calculated

1. Total tokens = response size / 4 (standard approximation)
2. Used tokens = size of only the fields your code accesses (detected via static analysis of data.fieldName, response["key"], etc.)
3. Wasted tokens = total - used
4. Cost = tokens * model pricing (GPT-4o: $2.50/M, Claude Sonnet: $3.00/M, etc.)

Fix It

The waste llm-tax reveals is exactly what SelfHeal fixes.

Send a target_schema with your API calls through SelfHeal's proxy, and it normalizes bloated responses down to exactly the fields you need — using LLM-powered transformation with x402 outcome-based pricing.

• Already compliant? Free.
• Needs normalization? $0.001 USDC. Only charged when it delivers value.

curl -X POST https://selfheal.dev/api/x402/proxy \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://api.example.com/users",
    "method": "GET",
    "target_schema": {
      "type": "object",
      "properties": {
        "id": { "type": "number" },
        "name": { "type": "string" },
        "email": { "type": "string" }
      }
    }
  }'

47 fields in, 3 fields out. Token savings: ~85%.

Privacy

• llm-tax never sends your code anywhere
• It only hits the API URLs it finds in your code (GET requests with no auth)
• No data is uploaded, tracked, or stored
• The --dry-run flag lets you see what it found without hitting any APIs

License

MIT

Built by Freedom Engineers