agent-landing

Can an AI agent actually use your product? And did your last launch land with the agents, or only with the people?

Agents now read your docs, call your API, and ship code against it, often before anyone on the team opens your landing page. Most product sites are still built for human eyes. They have pages that render only in JavaScript, no llms.txt, no markdown versions of the docs, and an API spec an agent can't find. Most launch dashboards still count human clicks too, so they can't tell when agents adopted a feature or kept retrying it until they gave up.

agent-landing checks both. scan shows you what an agent sees when it arrives and ranks the fixes by points. score reads your own launch data and gives a verdict for each audience you care about.

I scanned my own sites first

I built the framework behind this tool, then pointed scan at my own demo site. It scored 40 out of 100. The pages were readable without JavaScript and an agent could get in without a login wall, but there was no llms.txt, no markdown, no sitemap, and no API spec. The top three fixes on the list were worth up to 35 more points. So I worked the list and added explicit AI agent rules in robots.txt, a sitemap, JSON-LD, an llms.txt, markdown versions of the page, a read-only API over the demo data, and a read-only MCP server. The rescan scored 100.

My personal site, andyrosic.com, scored 53, an F. I worked the same fix list: explicit AI agent rules in robots.txt, an llms.txt, markdown versions of the pages, a read-only API with an OpenAPI spec, and a read-only MCP server. The rescan scored 100. Every one of those fixes came straight off the report, and the before and after is public.

score catches the opposite problem. In the bundled example, an auth refactor scores 76 under the human lens, which looks like a win. Under the agent lens it scores 35, and the tool flags why:

auth-refactor
  Not yet. It looks landed under the Human-adoption lens, but it is at risk under the Agent-as-user lens.
  Human-adoption lens   76/100  Landed
  Agent-as-user lens    35/100  At risk
  alert Retry rate spiked to 11.7%. A human-adoption dashboard would have called this a win and missed the failure.

The AI Hypocrisy Highlighter

Then I ran scan on the homepages of 52 companies that sell AI agents. The class average is 61, a D, and 22 of them got an F. Four turned the scanner away at the front door with a 403. Every company has a report card with a letter grade, a teacher's comment, and its cheapest fixes at hypocrisy-highlighter.vercel.app.

Why use it

• You define what landed means. Signals, weights, thresholds, lenses, and verdict labels live in one config file. Add a partner lens, drop the API checks for a marketing site, or weight retry rate twice as heavily as clicks.
• Every fix has a number on it. The scan ranks its fixes by how many points each one is worth, so the first item on the list is the one to do first.
• Claude sets it up for you. The Claude Code skill asks who your agent users are and what data you have, then writes the config and walks you through the results.
• It runs locally and costs nothing. No LLM calls and no paid APIs. The scanner respects robots.txt, caps page sizes, and won't fetch private network addresses unless you ask it to.

What it checks

It answers two questions for any product:

1. Can an agent find and use this product? agent-landing scan <url> checks the surfaces agents actually read (robots.txt, llms.txt, sitemaps, server-rendered HTML, markdown, JSON-LD, API specs, MCP discovery, auth friction, and token cost) and returns a score per check, an overall score, and a prioritized fix list.
2. Did a launch land for agent users? agent-landing score --data <events> scores each feature from your own telemetry under the lenses you define (human, agent, partner, or any other), with verdicts and a plain-language explanation such as a retry spike callout.

Signals, weights, thresholds, lenses, verdict labels, and scan settings all live in one validated config file. The CLI makes no LLM calls and uses no paid APIs.

The repo also holds the original demo site, live at agent-landing-rosic.vercel.app, which shows the framework on synthetic Vercel data.

Quickstart: CLI

Run it with npx, no install needed:

npx agent-landing scan https://your-product.com

Or install it globally with npm install -g agent-landing. Then, in your own project:

agent-landing init                               # writes agent-landing.config.json
agent-landing scan https://your-product.com      # terminal report
agent-landing scan https://your-product.com --md --out scan.md
agent-landing score --data launches.csv --json
agent-landing score --example                    # synthetic demo data

Option	Applies to	What it does
--config <path>	scan, score	Config file. Defaults to ./agent-landing.config.json when present, else built-in defaults.
--json, --md	scan, score	JSON or Markdown output instead of the terminal report.
--out <path>	all	Write output to a file. For init, the config path.
--pages /,/docs	scan	Override the key pages for one run.
--allow-private	scan	Allow loopback and private network targets. For scanning your own local dev server only.
--force	init	Overwrite an existing config.

Exit codes: 0 on success, 1 on a runtime or config error, 2 on a usage error.

This is the beginning of the README. Read the rest at the original. Read the full README ↗