· 8 min read

Chrome's Modern Web Guidance Gives Coding Agents 250 Web Features and Baseline Data in One Install. It Is a Preview With Telemetry On. Here Is How I'd Try It.

Stack Tools AI

Ask a coding agent for a modal and you will often get a div, a focus trap written by hand and a scroll lock, when the dialog element has handled most of that for years. A repo under the GoogleChrome organization, called Modern Web Guidance, is built to fix that habit. It hit the Hacker News front page on November 10 under a Show HN title about getting your agent off roughly 2019 JavaScript. The README says it covers 250 modern web features and 203 real-world use cases. I think it is worth trying this week, with two caveats about telemetry and maturity.

What it is

The README describes it as a set of skills that embed web platform expertise, best practices and browser compatibility data into coding agents. In practice it is a small skill plus a command line tool. The agent calls search with a plain-language query, gets back matching guide IDs, and calls retrieve to load one guide. The skill itself is small, around 200 tokens by the README's count, so it does not eat your context until the agent needs a guide.

The feature guides split into CSS and layout (140), HTML and DOM (36) and JavaScript and APIs (74). Use cases are grouped under headings like accessibility, forms, performance, security and UI components. There is also a second pack for Chrome extensions covering Manifest V3.

Compatibility data comes from the Baseline project, per the README, and the adoption rules lean on progressive enhancement with small fallbacks. Features in origin trial or still volatile are searchable but blocked on retrieve unless you turn them on with a flag, which writes to a .mwgrc file.

The license is Apache 2.0. The README calls it a preview release, and the repo has no tagged versions yet.

How you install it

The quick start is an interactive wizard:

npx modern-web-guidance@latest install

You can try it without installing anything into an agent:

npx modern-web-guidance@latest search "animate a dialog modal backdrop"
npx modern-web-guidance@latest retrieve "animate-to-from-top-layer"
npx modern-web-guidance@latest legacy-scan .

For Claude Code and GitHub Copilot CLI, the README gives the same two plugin commands:

/plugin marketplace add GoogleChrome/modern-web-guidance
/plugin install modern-web-guidance@googlechrome

It also lists install lines for Google Antigravity, Grok Build CLI and Kimi Code, plus the Vercel Skills CLI and the GitHub CLI as channels. One caution about what I read: the repo page and the raw README I fetched were different snapshots. They disagreed on whether Codex is in the install list, on the feature counts (the repo page showed 103 features and 131 use cases) and on the wording of who stewards the project. I have used the raw README where they differ. I did not see a Cursor install line in either, though the repo tree contains a .cursor-plugin directory.

The telemetry default

Telemetry is on unless you turn it off. The README says the tool collects installation counts, the name of the calling agent, the IDs of guides retrieved and the search queries the agent generates. It says raw user prompts are not collected, and the data is handled under Google's privacy policy.

The search queries are the part to think about. An agent writes them from your task, so a query can carry a hint of what you are building even if your prompt never leaves your machine. For a client project under NDA I would opt out before the first run:

export DISABLE_TELEMETRY=1

The README also says search runs locally with an offline embedding model and BM25, with no API keys. The repo page snapshot described the search implementation differently, so I would confirm that by watching network traffic on the first run if it matters to you.

What the evals say

The project publishes an eval dashboard, and the README's table for November 4 to 8 reports Claude Code on Opus 5.5 scoring about 56 to 57% unguided and 88 to 90% with the guidance. Those are the project's own numbers on the project's own tasks. They measure whether the agent used the modern feature the eval was looking for. They do not tell you whether your users' browsers support it or whether the result was better for your product.

The README is candid that the guidance adds little when a task has low "opportunity," meaning there was no modern feature to reach for.

What I'd actually do

  1. Run legacy-scan on one repo first. It costs nothing and touches no agent config. If it finds little, you have learned your code is already current and you can stop.
  2. Set DISABLE_TELEMETRY=1 in your shell profile if any of your work is confidential. Then decide later whether to send data back.
  3. Install it into one agent in one project, not globally. I would pick the project with the most UI code.
  4. Give the agent a task you have done before by hand, such as a dialog, a popover menu or a form with validation. Compare the diff with and without the skill. Ten minutes of reading tells you more than the dashboard.
  5. Check the output against your own audience. Baseline tells you what current browsers support. If your analytics show a slice of users on old Safari or embedded webviews, keep the fallbacks the guidance suggests and test them.
  6. Leave origin trial and volatile features off. They are off by default for a reason, and a solo product does not need to ship on an API that may change.

Where this could be wrong

A preview from a browser vendor can stall. With no tagged releases and @latest in every command, the guidance your agent reads can change between runs without telling you. If that bothers you, pin a version once one exists.

There is also a fair argument that this is a browser maker steering agents toward its own roadmap. The stewardship line in the README names the Chrome, Edge, Firefox and MDN teams, but the repo page I saw named only Chrome and Edge alongside the community, and I could not tell which is current. Until that is clear, I would treat the Baseline gating as the safeguard and not the list of names.

Even so, the problem it targets is real. Agents learned the web from years of old tutorials, and I have deleted enough hand-rolled modals to want this to work.

Author

Lukas

Sources

More Stack articles →

Newsletter

The week, in one email.

The weekly summary of what happened and what it means for people who build alone. No spam, unsubscribe any time.

Form not showing? Subscribe on this page instead.

Related Posts

AWS Shipped a Free Local Emulator That AI Agents Keep Up to Date. Telemetry Is On by Default and IAM Is Not Checked. Here Is How I'd Use It.

AWS Emulator runs AWS APIs on your laptop with no account and no fee, and it hit about 785 points on Hacker News. It also skips IAM policy checks, keeps state in memory, and sends usage data to AWS unless you switch it off. Good for fast tests, wrong for security tests.

Apple's Claris Is Testing Harbor, a Private AI Server That Pools Memory Across Macs Over Thunderbolt 5. Here Is What It Would Replace in My Setup.

Claris, an Apple company, has a private preview page for Harbor: run open-weight models on Macs you own, and pool unified memory across two or more Thunderbolt 5 Macs. No price, no named models, no API spec. Here is what is confirmed, what is not, and the math to run before you buy hardware for it.

Google's Gemini Decisions API Scores Up to 1,000 Questions per Request at $0.10 per Million Input Tokens. It Changes Yesterday's Self-Host Math.

Google launched a hosted decision model on November 11 that returns probabilities instead of text, bills only input tokens, and lets one request answer up to 1,000 questions. Against a rented GPU, the self-host break-even jumps from about 1,500 calls a day to tens of thousands. Test the accuracy before you switch.