Chrome DevTools for coding agents
| Rating | Summary | |
|---|---|---|
| Quality | excellent (23/24) | Actively maintained by Google, exhaustively documented with per-client setup guides, full test infrastructure, zero runtime deps (bundled), and explosive adoption. |
| PAI Relevance | integrate (0.88) | Fills a genuine DevTools-level gap in PAI's browser stack — performance tracing, source-mapped errors, and network inspection are absent from existing Browser and Interceptor skills. |
23/24 — actively-maintained / well-documented / high-discipline
Failed:
.github/workflows/ reference visible in the provided README content; CI presence is implied by the professional setup but cannot be confirmed from available evidence.Passed:
chrome-devtools-mcp-v1.1.1 tagged 2026-05-27.archived: false.Failed:
Passed:
npx JSON config and per-client installation guides for Amp, Antigravity, Claude Code, Cline, Codex, Copilot, etc.mcpServers config, --slim, --headless, and --browser-url usage.docs/tool-reference.md; "MCP Client configuration" section covers configuration options.docs/cli.md, docs/tool-reference.md, docs/slim-tool-reference.md, docs/troubleshooting.md, docs/design-principles.md, and CHANGELOG.Failed:
Passed:
package.json present and fully specified.dependencies listed; all 34 entries are devDependencies bundled via Rollup into a self-contained CLI artifact."test": "npm run build && node scripts/test.mjs" in package.json; sinon in devDependencies; separate test:only and test:update-snapshots scripts indicate structured test infrastructure.| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 2 | The combination of DevTools Protocol + Puppeteer wrapped as a structured MCP tool surface — with slim/full mode splits, per-tool telemetry, and performance tracing via Lighthouse + CrUX — offers directly applicable architectural patterns for upgrading PAI's Browser and Interceptor skills to support debugging and profiling workflows, not just navigation and scraping. |
| Integration Readiness | 2 | TypeScript package, npx -y chrome-devtools-mcp@latest or bun x invocation, MCP JSON protocol output, CLI binary entry points, zero runtime dependencies, Node LTS compatible — fully aligned with PAI's CLI-first Bun-based stack; drop-in as a subprocess or MCP skill. |
| Overlap Risk | 1 | Partial overlap with PAI's existing Interceptor skill ("stealth Chrome automation") and Browser skill ("batch headless scraping") — both cover Chrome control — but neither approaches DevTools-Protocol-level inspection, performance tracing, or source-mapped stack traces. |
| Gap Fill | 2 | PAI has no existing capability for DevTools-level browser debugging: Lighthouse performance profiles, network request analysis, console error collection with source-mapped traces, or CrUX field data integration are absent from all 52 skills and 79 tools in the manifest. |
Composite: 0.88
Conservancy development sessions (SvelteKit, Cloudflare Pages): Add chrome-devtools-mcp to the Conservancy project root via .mcp.json (npx @chrome-devtools/mcp@latest --slim), then open a Claude Code session alongside bun dev. When iterating on components, the agent navigates to localhost:5173, reads source-mapped console errors, and inspects computed DOM state directly — eliminating the manual "describe what you see in the browser" relay that breaks autonomous debugging loops and forces context-switching back to the developer.
Petites Fugues performance advisory: From a Claude Code session, point chrome-devtools-mcp's Lighthouse tool at the Petites Fugues storefront URL. The tool returns structured JSON combining CrUX real-user field data with lab scores; an agent parses this and produces a ranked remediation brief (image sizing, render-blocking scripts, Core Web Vitals failures) — replacing the manual PageSpeed Insights → copy → summarize workflow that advisory reporting currently requires with a reproducible, agent-generated deliverable.
Conservancy post-deploy regression checks: After each Cloudflare Pages deploy, run a Lighthouse audit via chrome-devtools-mcp against the production URL in a Claude Code session and have the agent diff the result against a lighthouse-baseline.json committed to the repo. Any score regression surfaces immediately as a structured delta — replacing the ambient uncertainty of "I wonder if that layout change hurt performance" with a deterministic, automated check that runs in the same session as the deploy.
Category: Web & Browser Automation
In this category: unclecode--crawl4ai (solid 18/24, integrate)
Standing: Significantly outperforms crawl4ai on quality (23 vs 18) and occupies a complementary niche — crawl4ai is LLM-optimised web scraping while chrome-devtools-mcp is agent-driven browser debugging and performance analysis; together they cover the full web automation spectrum in PAI's stack.
Density: 10/10 — Full README (8KB), complete dependency manifest (package.json with scripts, devDeps, engines), repo metadata (stars, forks, topics, dates, license, release), and GitHub org context (ChromeDevTools/Google LLC authorship) all present. Missing: source tree listing, .github/workflows/ YAML files, CHANGELOG content, and test suite files — none of which materially affect scoring at this evidence density.
This is the first official Google-authored MCP server for browser tooling, which distinguishes it structurally from community MCP browser projects. The --slim mode flag is a notable design decision — it limits the tool surface exposed to agents, reducing token overhead and accidental scope. The built-in usage statistics collection (on by default, opt-out via flag) is a Google-standard telemetry pattern; PAI integrations should include --no-usage-statistics in invocation config. The --browser-url flag enabling attachment to an already-running Chrome instance (used in Antigravity's config) is the most relevant integration path for PAI's headless agent workflows.