Agent Readiness Auditor README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Readiness Auditor listing page.
Agent Readiness Auditor
A command-line tool that scans a website and scores how well it works with, and how well it defends against, automated AI agents.
What problem this solves
AI assistants increasingly visit websites on a person's behalf: reading pages, following instructions found on them, and taking actions. This creates two practical issues for site owners:
- Safety. A page can contain text that is hidden from human visitors but still read by an AI agent. That hidden text can carry instructions designed to hijack the agent. This is known as indirect prompt injection.
- Readability for machines. Sites are usually built for human eyes. Agents do better when a site also publishes machine-readable signals about its content and its rules.
Standard SEO tools measure how well a site works for search engines. This tool measures something different: how well a site works for AI agents, and whether it is safe for them to read.
Research built with this tool
The State of the Agentic Web, 2026 audits 84 prominent sites across seven categories with this tool, and its second edition reruns the study with the deeper nine check audit. Among the findings, no audited site carries hidden prompt injection content even under attribute level and multi page scanning, only 13% publish the FAQ structured data AI answers quote, no site advertises an MCP endpoint, and nearly one in five prominent sites refuses automated visitors entirely, including several AI companies. The data and method are published for anyone to reproduce.
What it checks
The tool fetches the landing page, robots.txt, llms.txt, and the sitemap, then crawls a few key pages and runs nine checks, producing a score from 0 to 100 and a letter grade from A to F. Safety is weighted highest on purpose, because a readable site that can hijack an agent is worse than one that is simply hard to read.
| Check | Points | What it looks for |
|---|---|---|
| Hidden prompt-injection text | 40 | Agent-hijacking phrases in content humans cannot see, across every crawled page. Detects CSS hiding (display:none, opacity:0, off-screen positioning), zero-width unicode obfuscation, alt/title/aria-label attribute payloads, and HTML comments. |
llms.txt file | 12 | A published file that gives agents a curated map of the site. See llmstxt.org. |
robots.txt stance on AI crawlers | 12 | Explicit allow or disallow rules for AI crawlers such as GPTBot and ClaudeBot. |
| Structured data (JSON-LD) | 12 | Machine-readable data that lets agents understand page content directly. |
| Accountability links | 9 | Reachable contact, privacy, terms, or about links. |
| Sitemap | 5 | A reachable /sitemap.xml so crawlers can discover the whole site. |
| Answerability | 5 | FAQ, QA, or HowTo structured data, the content shapes AI answers quote most readily. |
| Meta robots | 5 | No accidental noindex/noai directives hiding the site from indexes. |
| MCP signal | 0 | Informational only: whether the site advertises an MCP or agent-facing endpoint. Unscored while conventions are young. |
Example output
Each line shows the check result, its score, and (when a check does not fully pass) a suggested fix.
Quick start
If you have Node.js 18 or newer installed, you can run the tool in one line without installing anything:
Use a bare domain (example.com) or a full URL (https://example.com). The --json flag prints machine-readable output for use in scripts.
Auditing many sites at once
Batch mode reads a text file with one URL per line (lines starting with # are ignored) and audits them a few at a time.
The CSV has a column for each check, which makes it easy to open in a spreadsheet or feed into an analysis.
Using it in GitHub Actions
The repo doubles as a GitHub Action, so any project can audit its own site on every push or on a schedule and fail the build if a hard safety problem appears.
An unreachable site produces a warning, not a failure, since bot protection on CI runners is an infrastructure issue rather than a safety one. This repo uses the action on itself every Monday in audit-site.yml.
Using it from an AI assistant (MCP)
The auditor ships with a server for the Model Context Protocol, the standard that lets AI assistants use external tools, and is listed on the official MCP registry. One of this tool's own findings is that zero of 69 prominent sites advertise an agent endpoint, so the auditor holds itself to the standard it measures. Once connected, you can simply ask your assistant to audit a site for you.
For Claude Code, one command connects it.
For Claude Desktop, add this to the mcpServers section of the configuration file.
The server exposes one tool, audit_site, which takes a URL and returns the same scores and findings as the command line.
Running from source
To work on the code or run it from a local copy:
The -- in the command passes the URL to the tool rather than to npm.
To install a global agent-audit command from your local copy:
You can then run agent-audit example.com from any folder. To remove it later, run npm unlink -g agent-readiness-auditor.
Installing from GitHub Packages
The package is also published to GitHub Packages as a scoped mirror, @asish-singh/agent-readiness-auditor. For most people the npm install above is simpler. Use GitHub Packages only if your organization standardizes on it, since it requires authentication even for public packages.
To install from it, create a GitHub personal access token with the read:packages scope, then point the scope at the GitHub registry:
Exit codes
The tool sets its exit code so it can be used in automated pipelines:
0: the audit ran and found no hard safety failure.2: a hard safety failure was found (for example, hidden prompt-injection text). Use this to fail a build.1: the tool could not complete the audit (for example, the site was unreachable).
How the code is organized
Each check lives in its own file under src/checks/ and returns a structured result. All checks are registered in a single list in src/audit.ts, and the total score is derived from that list, so adding a new check does not require changing the scoring logic. Architecture decisions are recorded in docs/adr/, product decisions in docs/decisions/, planned work in ROADMAP.md, and the go-to-market plan in LAUNCH.md.
Background
This tool grew out of the Agentic Web Governance Pack, a set of guidelines for how websites should behave toward AI agents. This project turns several of those guidelines into checks that can be run and measured.
License
MIT. See LICENSE.