Chrome ships Modern Web Guidance 1.0 to fix coding agents' old habits
Google’s Chrome team released Modern Web Guidance 1.0 on November 10, 2026: an open-source set of agent skills and a local CLI that feed coding agents expert guides on modern HTML, CSS and JavaScript features, so they stop writing old <div>-and-JavaScript patterns for things the browser now does natively. Firefox, Edge and MDN now help maintain it, and it works with Claude Code, Codex, Copilot CLI, Antigravity and other agents.
What’s new
Chrome showed an early preview of Modern Web Guidance at Google I/O in May. The 1.0 announcement on the Chrome for Developers blog lists four changes for the stable release:
- A modernization workflow that scans an existing codebase (HTML, CSS, JavaScript and
package.json) for outdated patterns and guides the agent through upgrading them. For now it uses static analysis on a limited set of patterns; Chrome says it plans to widen it. - A bigger guide database aimed at what Chrome calls “verified AI knowledge gaps”, with cross-browser solutions.
- Human expert review for every guide, including a dedicated accessibility review, with user-facing guides tested on real assistive technologies.
- Evals for every guide, published on a public eval dashboard.
Firefox and Edge have joined Chrome on the project, and the GitHub repository now describes it as stewarded by the Chrome, Edge, Firefox and MDN teams. The README lists 250 web features covered: 140 in CSS and layout, 36 in HTML and DOM and 74 in JavaScript and APIs, from :has() and container queries to View Transitions, the Popover API, Anchor Positioning, scheduler.yield and the browser’s built-in AI APIs. There is also a separate skill pack for Chrome extensions (Manifest V3, service workers, Web Store publishing).
Paul Irish of the Chrome team, who posted the release on Hacker News, gave examples of what models get wrong: CSS.highlights for highlighting text ranges, <dialog> with showModal() for backdrops and focus handling, and field-sizing to auto-grow textareas. Models “know about some of these by name”, he wrote, but still default to the older pattern.
How it works
The skill tells the agent to call the modern-web-guidance CLI for web platform questions. The agent runs search with a plain-language query, the CLI finds the best guide with an offline embedding model and BM25 search, and retrieve puts the guide’s code patterns, gotchas and fallbacks into the agent’s context. According to the README, search runs locally with no network calls or API keys, and the npm package has no dependencies.
Guides carry Baseline browser support data, so the agent can pick a feature that fits a project’s support targets and add small fallbacks instead of heavy polyfills. Guides for Origin Trial APIs and “volatile” single-engine features are gated: an agent can find them, but only retrieves them when the project opts in through a .mwgrc file.
What it means for developers
Install it from a project directory with the wizard:
npx modern-web-guidance@latest installProjects on an earlier preview update with npx modern-web-guidance@latest update. It can also be added with npx skills add GoogleChrome/modern-web-guidance, as a Claude Code, Copilot CLI or Codex plugin, or with Antigravity, Grok Build and Kimi Code; the README has the command for each. To try it without installing:
npx modern-web-guidance@latest search "animate a dialog modal backdrop"
npx modern-web-guidance@latest legacy-scan .Two things to know before adding it to a team’s setup. The CLI sends Google anonymous usage data, including the calling agent’s name, guide IDs and the search queries the agent writes (not your prompts); set DISABLE_TELEMETRY=1 to turn it off. And the code is Apache 2.0 while the guide content is CC BY 4.0, partly derived from MDN.
The guides are also readable as plain documentation, which Chrome says many developers use to learn new features themselves. If your stylesheets still look like our old roundup of ways to center an element in CSS, the legacy scan is a quick way to see what the platform can now do for you.
How it compares
The repository’s eval snapshot for November 8 shows each agent passing more of its roughly 215 web development tasks with the guidance than without it:
| Agent (model) | Without guidance | With guidance |
|---|---|---|
| Antigravity CLI (Gemini 4 Argon) | 58% | 94% |
| Claude Code (Opus 5.5) | 57% | 88% |
| Codex CLI (GPT-6.1 Sol) | 54% | 82% |
The tasks are Chrome’s own, run on its sample apps and graded by Playwright checks, so treat them as the project’s numbers rather than an independent benchmark.