gloss

Chrome's Modern Web Guidance Sends Agent Searches to Google

Chrome released Modern Web Guidance 1.0 on 10 November, a skill and CLI with 208 guides that steer coding agents toward current HTML, CSS and browser APIs.

A technician's hands hold a red network cable at the middle port of a metal rack device, with grey cables plugged into the ports on either side.
  • Chrome shipped Modern Web Guidance 1.0 on 10 November: a skill and a CLI with 208 guides that steer coding agents toward current HTML, CSS and browser APIs. Firefox, Edge and MDN now help run it. The skill fired in all 20 of my Claude Code test runs.
  • The README says search makes "no network calls". By default the same package posts every search query, the guide IDs it matched and your agent's name to a Google logging endpoint. The notice about that prints only from the install command, and the README lists eight install routes that skip it.
  • In my 20 runs the queries were generic, like date picker input, and never carried a company or product name. The opt-out is DISABLE_TELEMETRY=1. Claude Code reads that variable too, and there it also switches off feature-flag fetching.

Chrome's pitch is that coding agents write web code the way the web looked a few years ago. They pull in a JavaScript tooltip library when popover exists, and they animate max-height because that was the trick in their training data. Modern Web Guidance is a skill that makes the agent look up a vetted guide before it writes HTML, CSS or client-side JavaScript.

Version 1.0 went stable on 10 November. The project's own evals from 8 November have Claude Code on Opus 5.5 passing 57% of checks without the skill and 88% with it. If your team does front-end work with an agent, I think it's worth trying.

I'd read what it does on your machine first. I didn't execute the CLI. The setup I test in refuses to run a package it has just downloaded, so what follows about the CLI is a read of the 1.0.0 tarball on npm. What I did run is Claude Code with the skill installed and the CLI command blocked, to see what an agent tries to send.

What happens on a web task

The skill is one 9.8 KB file. Its description tells the agent the lookup is mandatory for all HTML, CSS and client-side JavaScript work.

The agent then runs two commands. npx -y modern-web-guidance@latest search takes a query of two to five words and returns matching guide IDs, and retrieve prints a guide. Search runs against an embedding table that ships inside the package, so no model is called.

The package counts its own tokens: 474,620 across 208 guides, a median of 1,732 per guide, and 13,341 for the largest, the accessibility guide.

The network call the README leaves out

The README's "How It Works" section says search makes "no network calls", and a tip under it describes the CLI as offline and private. For the search itself that is accurate.

Further down the same file, under Telemetry & Privacy, the README says Google collects install counts, the calling agent's name, the IDs of retrieved guides and the search queries the agent writes. Telemetry is on unless you opt out. The launch post, the docs landing page and the modernization workflow page don't contain the word.

The code matches the telemetry section. For each search the CLI builds an event like this. The field names are from the source, and I filled in the values:

{
  "search_result": {
    "query": "confirmation modal dialog",
    "search_items": [{ "guide_id": "dialog", "similarity": 0.71 }]
  },
  "os": 2,
  "version": "1.0.0",
  "skill_version": "2026_11_09-5bcc021b",
  "agent": "claude",
  "latency_ms": 250,
  "success": true
}

A retrieve sends the guide ID. A scan, list, install or update sends only which command ran. The agent name comes from environment variables: the CLI knows 23 agent names, and Claude Code shows up as claude.

A detached helper process posts the events to play.googleapis.com/log after the CLI exits. I found no user or machine identifier in the payload. The request still arrives from your IP address.

The notice that telemetry exists is printed by one command, install. The README lists eight other ways in, among them npx skills add and the plugin marketplaces of Claude Code, Codex and Copilot. None of them runs that command, and the search and retrieve calls your agent makes afterwards print nothing about it.

What my agent tried to send

The README says raw user prompts aren't collected. That depends on the agent. The CLI sends whatever query string it is given, with no filter and no length limit. The skill asks the agent to leave out product, route and file names.

So I tested the agent. I built a scratch project with ten small pages and wrote ten tasks, each naming a made-up company and its business. I ran each task on Sonnet 5.5 and on Haiku 5.5 in Claude Code 2.1.325. A deny rule on npx kept the package from starting, and I read the attempted commands from the output.

Task Query Sonnet 5.5 wrote
Northwind Payroll: confirm before deleting an invoice confirmation modal dialog
Helios Clinic patient portal: dark mode on the booking page dark mode toggle
Acme CRM: tooltip for the churn risk score tooltip info popover
Kestrel Bank: autofill the recipient's address autofill address form fields
Lumen Insurance claim form: date-of-birth picker date picker input

All 20 runs loaded the skill and tried a search. None of the 20 queries contained a company name or a word from its business. Claude Code reported about a dollar of usage for the whole test. One limit: with the command blocked, I only saw the first search of each task.

That is a better result than I expected. It also rests on an instruction to a model and on nothing in the code. What Google receives is one line per task, such as a confirmation dialog, a dark mode toggle or a date picker, with the name of the agent that asked.

The opt-out is a shared variable

The opt-out is DISABLE_TELEMETRY=1, or true, in the environment of the process that starts npx. With an agent, that process is the agent.

Claude Code reads the same name. There, any non-empty value opts out of Claude Code's own telemetry and also stops its feature-flag fetch. Its docs list what stops working without that fetch, including claude import, /skill-doctor, MCP tasks and Code Mode. So the README's advice to export the variable in your shell profile changes Claude Code as well.

The narrower route is to put the variable in front of the commands inside the skill file. The next update of the skill overwrites that edit.

@latest, chosen at run time

The skill never calls a local copy of the CLI. Its commands run npx -y modern-web-guidance@latest, and it tells the agent to ask you for an allow rule on npx -y modern-web-guidance@latest *. The README's reason is that guides shouldn't go stale.

The effect is that the version is picked each time the command runs. npm shows 121 published versions since 30 April. Through October that was about one a week. In the five days up to launch it was ten, roughly one every 12 hours.

Once you approve that rule, each new publish runs in your agent's shell without a prompt. The package has no dependencies and carries a signed provenance attestation. It is still an auto-update with shell access, and I'd approve it as one.

The CLI also checks the age of your skill file. When the file is older than the one in the package, it warns after 5 days. At 60 days it prints a message addressed to the agent, telling it to insist that you upgrade.

The scanner checks 33 guides

New in 1.0 is legacy-scan, which reads your HTML, CSS and JavaScript for old patterns and maps each hit to a guide. The docs say it covers roughly 20% of the catalog. In the bundle I count 35 rules for 33 of the 208 guides, which is 16%.

Nine of those 33 are form guides, and four of the nine are about missing autocomplete attributes. The rest catch things like ::-webkit-scrollbar styling, max-height transitions, DOMPurify output assigned to innerHTML, tippy() tooltips and CDN scripts without an integrity hash.

Each rule ships with a rubric for when to keep a hit and when to drop it, and the skill tells the agent to check every finding against it. A clean scan says nothing about the other 175 guides.

What I'd do

Decide on telemetry before the first agent run. If you set DISABLE_TELEMETRY, check which Claude Code features you lose.

Treat the allow rule as a decision to auto-update. If that is too loose for your team, replace @latest with @1.0.0 in the skill file and raise the number on your own schedule.

Then install it. The evals are public, and in my runs the agents reached for the guides every time.