Modern Web Guidance
Modern Web Guidance is a knowledge pack for coding agents. It gives them guides for modern web work before they write code, so they work with current practice on accessibility, forms, performance, security, and user experience.

Two words first
What matters here first is the intent. You are not getting a library to import. You are getting material that changes what the agent thinks about before it writes code. You can see that from the start: the guide does not throw generalities at you, but sets rules that touch the real problem, from whether a skip link exists to whether a modal traps focus correctly. That is where its value sits. It does not talk about quality in the abstract. It turns quality into small decisions that can be checked before code is written.
What matters second is the range. It does not speak to one narrow problem. It covers many parts of the web that shape how code is designed and written.
That makes it feel different from a normal package. Here, the package does not produce an app. It produces a frame of reference.
What it does
Modern Web Guidance loads guides that come before writing. The idea is simple. If the agent sees the right rules first, the odds of old or sloppy web code go down.
The main project repository is https://github.com/GoogleChrome/modern-web-guidance and the source code lives at https://github.com/GoogleChrome/modern-web-guidance-src. The npm package is called modern-web-guidance.
This is not one single instruction. It is a bundle of material. The install places SKILL.md files and guides for accessibility, css, css-layout, forms, html, js, performance, privacy, pwa, security, ui-atoms, ui-behaviors, ui-components, user-experience, visual-design, wasm, and webmcp.
The content inside one of those guides matters too. The accessibility guide, measured at 51,734 bytes, starts with two principles that set the tone for the whole package: accessibility is the floor, not the ceiling, and patterns are case-specific; no checklist replaces real testing, including testing with disabled people. That sounds obvious, but in practice it cuts off many mistakes before they become code. If the agent knows a checklist is not enough, it will look differently at a modal, a form, or keyboard navigation. And if the guide has sections for landmarks, headings, lists, tables, skip links, accessible names, document language, focus, alt text, forms, live regions, color, typography, motion, and modals, then it is not just saying “make it accessible.” It is telling you exactly where the experience breaks and how to rebuild it.
The same depth shows up in the 209 use cases. ime-safe-enter-submit solves submitting with Enter without breaking text composition for anyone typing with an IME. validate-input-after-interaction avoids an error message before the user has even touched the field. fluid-scaling ties font size and spacing to the container itself, not to fixed breakpoints. animate-to-intrinsic-sizes gives a smooth open and close animation at an element’s natural size, like a collapsible menu. autofill-payment-form builds a card form with the right fields so the browser can fill it automatically. webmcp exposes page functions to AI agents in a standard way. These are not broad rules of thumb. They are specific failures that get fixed before they are written.
The guide is not theory for the shelf. It is a toolbox that goes into detail. An agent that reads it before writing a page structure or a form will think differently about what counts as real control and what is only visual imitation.
The package is written for coding agents. A person does not read it like a reference book, and the browser does not use it like a dependency. The agent loads it, and it does so to change the way it decides.
What it does not do
It is not an MCP server. That matters, because it is easy to mistake it for a service that serves tools over the network. Here we are talking about a CLI that installs knowledge, not endpoints.
It is not an application library. You do not drop it into a project to build features with functions. You do not import APIs. It does not give you runtime behavior.
It does not replace the judgment of a person or an agent. If the agent does not know how to apply a guide well, the material alone will not save it. It gives a better starting point, not automated thought.
It is not limited to one framework, and that is both strength and limit. It is a strength because the material covers the whole web. It is a limit because it will not solve product-specific problems if the core architecture is already weak.
Our hands-on test
On 18/11/2026, we tested it in a clean folder. That is where you see whether a tool stands on its own feet or needs a ready-made environment to look good.
The command npx -y modern-web-guidance@latest install downloaded the repository, found 2 skills, installed 1 named modern-web-guidance, and left 2.5 MB in the .agents/skills/modern-web-guidance folder.
That same install wrote 17 guides and one SKILL.md file measuring 9,891 bytes. It also wrote a .claude/skills/ link and skills-lock.json. The lock pinned the install to a specific commit, 64a6d708…. That is a good sign. You know what you got.
What remains after the run is the habit of the tool. The 17 guides stay in .agents/skills/, SKILL.md stays as the 9,891-byte entry point, and the .claude/skills/ link shows that the same material can appear in a different client without being copied. skills-lock.json is the part that matters most. It locks everything to one specific commit. Without it, the same tool could behave differently on two machines on the same day. With it, you know you are getting the same material, not a moving version of it.
In the same pass it declared support for Claude Code, Codex CLI, Cursor, Antigravity, Gemini CLI, Amp, Cline, and others. The tool set up the universal skill for 21 agents and the link layer for Claude Code.
list returns 209 use cases. search "form validation" returns 1,983 bytes. retrieve accessibility returns a 51,734-byte guide. legacy-scan . in a folder without old code returns 0 findings. The same run also says something else: the legacy-pattern checks cover 38 guides, about 20% of the material. That is a limit. legacy-scan does not judge the whole package. It judges one fifth of it.
The numbers around the project are clear too. Downloads are 261,186 in the last week and 554,305 in the last month. The main repository has 22,335 stars and 889 forks, while the source repository has 8,559 stars. The project was created on 23/03/2026.
The latest npm release is 1.0.1 and it was published on 16/11/2026. The package requires Node >= 22. And the installer says that Google collects anonymous usage data to improve the tool, and that telemetry is fully disabled with DISABLE_TELEMETRY=1.
From the test, the strength is clear. The tool installs itself, carries many guides, and pins the version. It does not make a scene. It simply leaves material the agent can use.
Using it from the shell
If you look at it practically, the flow is simple. First you install. Then you ask. Then you fetch the guide you want. That is the whole point of the CLI here.
The basic install is this:
npx modern-web-guidance@latest install
If you want to search for a use case:
npx modern-web-guidance@latest search "form validation"
If you want to fetch a specific guide:
npx modern-web-guidance@latest retrieve accessibility
If you want to measure what is there before you go deeper:
npx modern-web-guidance@latest list
list returns 209 use cases. That gives you the scope at a glance before you choose a path.
If you want to check whether a pattern fits old code:
npx modern-web-guidance@latest legacy-scan .
legacy-scan . returns 0 findings when the folder has no legacy code. And it also tells you the limit: legacy-pattern checks cover 38 guides, about 20% of the material. It is not a full judgment of the package. It is a check of one part of it.
legacy-scan is most useful when you already have code and want to see where an old pattern smells. In an empty folder it does not say much. In a real codebase it shows whether the material actually lands.
Installation
What matters here is that installation is not a vague “run something.” It is a concrete step that writes skills to disk and pins them to a specific release.
The command is this:
npx modern-web-guidance@latest install
If you want to see it in a clean folder, start in an empty directory first. Then run the command. That is the simplest way to see what it really writes into the file system.
The package also states that it needs Node 22 or newer. That is not our estimate. It is the package’s own requirement. Anyone with an older Node does not get in cleanly.
The package also says it works with many agents. That matters because it does not lock you into one client. The same skill can become shared ground for different code-writing tools.
Where it helps you most
It helps when you are building your first site and do not yet know which mistakes repeat. A guide for accessibility or forms gives the agent exact thought patterns before it writes code. That means you do not have to discover from scratch every time that a label is missing or that the keyboard does not reach where it should.
It helps when a team wants its bots not to write in ten different styles. The material keeps a common line. If one agent normalizes a form and another regenerates the same UI in a different way, the result looks patched together. Here the shared knowledge cuts off the difference before it gets into the code.
It also helps when you are maintaining an old site and want to see where an old pattern is still hanging around. That is exactly where legacy-scan is useful. You do not want it to flatter you. You want it to show where behaviors or structures no longer hold up. In a legacy codebase, that direct view saves time.
Where to avoid it
Avoid it when you want direct feature execution rather than guidance. If you simply want to add a feature, this tool is not a shortcut. It is a frame.
Avoid it when you do not have a process for agents. If your way of working does not pass through an agent that reads skills before writing, the package loses its point.
Avoid it when you are running Node older than 22. The package asks for that explicitly, and there is no room for polishing that away.
Finally, avoid it if you do not care about discipline on the web. The material is broad, but it has an opinion. If you want to write however you like, you will find it strict.
What stays with me
What stays with me is that it is new, specific, and practical. It does not pretend to be something else. It does not pretend to be a framework. It does not pretend to be a server. It is a skill pack for coding agents.
I also keep the lock to a specific commit. That gives confidence. You are not looking at a moving target. You can see what you installed.
I keep the fact that it has real breadth. The 17 guides and 209 use cases show that this is not a decorative repo. It is material with weight and purpose.
I also keep the clear telemetry note. I do not like having to guess what a tool sends back. Here that was written out openly.
What it costs
In money: nothing. Apache-2.0 licence, no key, no account, no subscription, no second service. The install left 2.5 MB on disk — that is the whole footprint.
The context tax is not fixed — you pay per guide. It is not an MCP server, so it does not declare tools that sit in every conversation: the fixed tax is zero. The cost shows up the moment you ask for something. The central SKILL.md file is 9,891 bytes, about 2,500 tokens by estimate (bytes ÷ 4), and a guide like accessibility, at 51,734 bytes, is about 12,900 tokens. For comparison, a mid-sized MCP server with seven tools loads about 1,168 tokens into every conversation even when you never use it. Here you pay only for what you read.
Measurement and estimate, kept apart: the byte sizes are our own measurements from running the tool; the token conversion is an estimate (characters divided by four), not a measurement. The zero in money is not an estimate: it is the Apache-2.0 licence and the absence of an account, which we checked at the registry and the repository.
On telemetry: the installer itself states that Google collects anonymous usage data and that it is fully disabled with DISABLE_TELEMETRY=1. So the zero cost is not paid for with your data, unless you want it to be.
Sources:
https://github.com/GoogleChrome/modern-web-guidance
https://github.com/GoogleChrome/modern-web-guidance-src
https://www.npmjs.com/package/modern-web-guidance
https://github.com/GoogleChrome/modern-web-guidance/releases
https://github.com/GoogleChrome/modern-web-guidance#readme
https://docs.npmjs.com/cli/v11/using-npm/config#engine-strict
Εγκατάσταση
1 ΒΗΜΑ- npx modern-web-guidance@latest install