Asking Claude Code to "read the console errors on this page" used to mean opening DevTools myself, pasting a screenshot, or copying the messages by hand. With Google shipping Chrome DevTools for agents 1.0 as a stable release on May 22, 2026, that friction is finally going away.
I am an indie developer running the four Lab sites (Claude / Gemini / Antigravity / Rork) on Cloudflare Workers, deploying almost every day. I tried the tool from Claude Code on launch day. Both the MCP path and the CLI path work, but there are some gotchas worth knowing, plus a meaningful difference from the version bundled in Antigravity 2.0.
What Chrome DevTools for agents 1.0 is
In short, it is the official debugging tool for AI agents that drive Chrome. Per Forest Watch coverage, the stable release covers user-action replay, site quality auditing (Lighthouse), device and geolocation emulation, extension debugging, and even memory leak detection.
The space was crowded — Playwright MCP, Claude in Chrome, third-party Chrome CLIs — and Hatena Bookmark commenters echoed that the choice was hard. With Google's "official" badge on it, I expect Chrome DevTools for agents to become the default option in the Claude Code ecosystem fairly quickly.
Installing from Claude Code
Claude Code has built-in commands for managing MCP servers, so installation is one line.
# Register as MCP (recommended: tool calls stay organized)
claude mcp add chrome-devtools \
--command "npx" \
--args "-y" "chrome-devtools-mcp@latest"
# Or use the CLI directly
npm install -g chrome-devtools-cli
chrome-devtools --helpRestart Claude Code and /mcp will list chrome-devtools. /mcp chrome-devtools tools shows the tool surface area, so listing them first saves later confusion. Antigravity 2.0 users do not need to install anything — it is bundled by default.
MCP or CLI for daily use
Having tried both, MCP is the obvious choice if Claude Code is your daily driver, for three reasons.
- Claude Code receives the tool list automatically, so "audit the performance of this page" works in natural language
- Tool schemas are explicit, so the model rarely builds malformed arguments
- Responses come back in a format the model can directly consume (JSON)
The CLI shines in CI, where Lighthouse becomes a deploy gate. That work does not need an agent in the loop and is lighter as a plain npm script. For my four Lab sites I run Lighthouse from the CLI inside the Cloudflare Pages build, and use the MCP path for ad-hoc investigation during development.
Where it earned its keep in indie-developer work
When I add a new tool, I look for three places it pays off within the first few days. Here are the ones that stuck.
1. Verifying paywall behavior
Because Claude Lab sells memberships through Stripe Checkout, articles with premium: true rely on cookie checks. Asking Claude Code "open /articles/claude-code/xxx and tell me what shows up without the Premium cookie" makes Chrome DevTools for agents load the page and return a clean HTML view. I used to do this by opening an incognito window manually; verification time dropped to roughly a third.
2. Catching console errors after deploy
Right after shipping things like the bfcache fix (pageshow + event.persisted), I want to confirm nothing is throwing in the console on production. "Open production /membership and list only the console errors and warnings" gives me a tidy summary via the new tool. If anything looks off I can pivot to "show me the diff that fixes it" in the same conversation.
3. Diffing iPhone vs desktop layouts
For per-resolution checks like 6_5inch or 5_8inch, device emulation pays off. Setting emulateDevice to iPhone 15 Pro Max and then cycling responsive breakpoints surfaces the CSS media query boundaries fast. If you work on wallpaper apps with resolution-specific .htaccess redirects, this also helps verify what production URLs really return.
Gotchas and workarounds
A few small traps I hit after a few days of use.
- CAPTCHA and Cloudflare Turnstile still block it — Hatena commenters flagged this, and yes, bot-protected pages refuse to load. My four Lab sites have Bot Fight Mode dialed loose enough on production that I can hit them myself, but on client sites you will need IP allowlisting in advance.
- Extension state is invisible by default — extension debugging works, but it spins up a clean profile by default. If you want your everyday Chrome extensions present, pass
--user-data-dirat startup. - Memory leak detection takes time — running Performance monitor continuously and reading the result takes 30 to 60 seconds per investigation. Parallelism in CI needs to be tuned or you will burn budget fast.
How it compares to the Antigravity 2.0 bundle
For Antigravity 2.0 users, here are the practical differences versus the Claude Code route.
| Item | Claude Code + MCP | Antigravity 2.0 bundle |
|---|---|---|
| Install | claude mcp add manually | Available out of the box |
| Invocation | Via MCP protocol | Via Antigravity's internal Agent Manager |
| Config persistence | ~/.claude/mcp.json | Per workspace |
| Sharing across projects | Global config shares across all | Per-project configuration |
In my CLI-first workflow, having one global install reachable from every project is convenient. In Antigravity-as-editor scenarios, per-workspace config is actually safer because you can scope permissions per project. The right choice depends on how you work.
First command worth trying
The fastest way to get a feel for it is to run a Lighthouse audit. From Claude Code, try something like "fetch the Lighthouse scores for example.com and show me Performance and SEO in detail." Scores and suggestions come back as JSON and the model can analyze them in the same response. If you ship AdMob-monetized indie apps with marketing landings, this is a smart entry point for Core Web Vitals work.
Even with an official release, the space is crowded with similar tools. Lining up Playwright MCP, Claude in Chrome, and Chrome DevTools for agents and picking what fits your workflow is the practical move. For now I keep all three: Claude in Chrome for general browsing, Playwright MCP for E2E tests, Chrome DevTools for agents for debugging.
If you run into trouble installing it, I am writing a companion piece on the Antigravity 2.0 bundled experience over on Antigravity Lab — reading both should give you the full picture.