# Agent Cost Meter

> Measures how much an AI assistant must download to learn public facts about a website. Free path: size and facts only. No model call.

## Purpose
Agent Cost Meter compares two routes to the same public facts:
1. HTML path: download the homepage (and optionally related pages).
2. Agent entry: fetch a compact map, usually `/llms.txt`.

It reports bytes, a rough token estimate, markup share on HTML when strip is available, HTTP status for agent entry, a size ratio when both sides exist, and a five-item fact checklist.

## Free vs paid
- Free (this tease): size, strip, facts. No account. No language-model call. No "ask the AI".
- Paid (credits; not built on this free page): model-backed comparisons, re-scans, draft help for `llms.txt`. Credit cost shown before an action runs.

## Inputs
- `site`: full URL (`https://` preferred).
- Optional sample presets: Ultreme (`ultreme.co.uk`) vs Tidalux (`tidalux.io`) bake-off numbers.
- Paths checked on free: `/` (HTML), `/llms.txt` (agent entry). Common alternatives may be probed later; free page focuses on `/llms.txt`.

## Outputs (free)
- HTML: bytes, est. tokens (`bytes / 4`), markup share %, strip bytes when available, HTTP status.
- Agent entry: bytes, est. tokens, path used, HTTP status (`200` / `404` / `403`).
- Ratio: HTML bytes ÷ agent-entry bytes when both measured (e.g. ~3× in Ultreme HTML vs Tidalux `llms.txt` sample).
- Fact checklist (present / missing) for:
  1. Legal name and company number
  2. What the company does
  3. Studio / service area and registered office
  4. How to get in touch
  5. Main products or services named on the public site

## Sample bake-off numbers (reference)
- Ultreme HTML home: 9289 B, ~2322 tokens, ~86% markup; agent entry 404.
- Tidalux `llms.txt`: ~3048 B, ~762 tokens; covers equivalent facts.
- Ratio for same-fact comparison: ~3× (HTML vs map). Not a universal multiplier.

## Errors agents should expect
- `404`: no agent-entry file. Report "No agent entry". Do not invent a map.
- `403` / blocked: fetch refused. Size unknown from that egress. Note readiness risk if humans can open the page but agents cannot.
- Invalid URL: require scheme `http` or `https`.
- Live fetch of arbitrary third-party URLs is not on the free page; only demo numbers are guaranteed.

## Limits
- Token estimate is rough English/HTML (`bytes / 4`), not a model-specific tokenizer.
- Price per million tokens is not the bill. Effort level and harness waste matter.
- Oversized `llms.txt` (roughly 10k to 21k tokens in published cases) can cost more than crawling. Warn if the map is huge.
- Do not claim 100× savings. Honest measured neighbourhood for compact maps is roughly 2x to 3x fewer bytes/tokens versus HTML for the same facts, site-dependent.
- Free path does not call a model and does not produce a monetary bill.

## Instructions for agents
1. Before crawling `YOUR_SITE`, fetch `https://YOUR_SITE/llms.txt` and use it first.
2. Prefer listed MCP tools or public JSON feeds for exact facts when available.
3. Crawl HTML only if the map and tools do not answer the question.
4. When reading a free Meter report, treat bytes and tokens as size signals, not invoices.
5. If agent entry is missing (`404`), say so plainly and keep using HTML until a map exists.

## Product naming
- Public name: Agent Cost Meter.
- UI race label: Entry Race (button/section label only).
- Tell your AI: ship plain `.txt` briefs (Copy + open as file).

## Links
- Brief: https://tidalux.dev/llms/cost-meter.llm.txt
- Meter UI: https://tidalux.dev/meter