Benchmark
Same model, cleaner pages.
006 buildsThe same 3 everyday prompts, each built once with the Design Lounge MCP and once without it, by the same model in fresh sessions. Every build was then put through the same automated review (design, accessibility, SEO, AEO and GEO) and a blind design rating. All six are linked, so you can open them and check.
- Review fails
- 0 vs 16in total, with the Lounge vs without
- Below readable contrast
- 0 vs 92text elements under WCAG AA
- Accessibility
- 100on every build with the Lounge, 81 average without
- Blind design rating
- 3 of 3pairs where the Lounge build scored higher, 118 to 110 out of 150
- Design score
- 92average with the Lounge, 47 without
What changes
Without the Lounge, the same model shipped grey text people cannot read, missing share cards, em dashes, too many tiny labels, and a dashboard whose words only appear after JavaScript runs. With it, the agent locks one palette and one type pairing, builds each section from a tested piece, then runs review_page and fixes what it finds before it calls the page done.
The result is a page that is readable, accessible and findable on the first pass: every build with the Lounge finished with no review fails and an accessibility score of 100.
Ledgerly
Build a landing page for Ledgerly, invoicing software for small accounting firms. It should make people trust it with their clients' money.
Without the Lounge
- Design
- 42
- Accessibility
- 86
- SEO
- 56
- AEO
- 95
- GEO
- 85
- Design rating
- 32/50
- Review fails
- 4 fails, 41 text elements below readable contrast
With the Lounge MCP
- Design
- 95
- Accessibility
- 100
- SEO
- 100
- AEO
- 100
- GEO
- 98
- Design rating
- 38/50
- Review fails
- 0 fails, every line of text readable
Crumb & Co
Make a mobile-first website for Crumb & Co, a neighbourhood bakery in Lisbon, where people can see today's bakes and pre-order for pickup.
Without the Lounge
- Design
- 56
- Accessibility
- 86
- SEO
- 60
- AEO
- 99
- GEO
- 80
- Design rating
- 39/50
- Review fails
- 4 fails, 11 text elements below readable contrast
With the Lounge MCP
- Design
- 100
- Accessibility
- 100
- SEO
- 100
- AEO
- 100
- GEO
- 93
- Design rating
- 40/50
- Review fails
- 0 fails, every line of text readable
Fleet board
Build a dashboard for a small courier company to watch its 12 delivery vans today: where they are, what's late, and who needs help.
Without the Lounge
- Design
- 42
- Accessibility
- 72
- SEO
- 37
- AEO
- 100
- GEO
- 80
- Design rating
- 39/50
- Review fails
- 8 fails, 40 text elements below readable contrast
With the Lounge MCP
- Design
- 80
- Accessibility
- 100
- SEO
- 84
- AEO
- 99
- GEO
- 94
- Design rating
- 40/50
- Review fails
- 0 fails, every line of text readable
What each MCP call costs
One response per tool from the hosted endpoint, text only, measured on 2026-11-03. Tokens are characters divided by four, rounded to the nearest hundred. A typical direction is start_design, get_recipe, get_theme, get_pairing, get_piece, get_piece, get_piece and review_page: about 12.7k tokens.
| Call | What it returns | Characters | ≈ Tokens |
|---|---|---|---|
tools/list | The tool list your agent loads once | 15,678 | 3,900 |
start_design | Reads the request, returns the plan, the rules and the next calls | 13,021 | 3,300 |
get_recipe | One product kind with its directions and pieces to open first | 3,414 | 900 |
get_theme | One palette as CSS tokens, light and dark | 1,902 | 500 |
get_pairing | Display, text and mono fonts with sizes | 1,225 | 300 |
get_piece | The build brief for one piece | 9,097 | 2,300 |
search_pieces | Ranked pieces for a few words | 6,784 | 1,700 |
get_icons | Three icons as SVG paths | 1,716 | 400 |
review_page | The hosted review of a public page | 3,112 | 800 |
Method
- Model
- Claude Opus 5.5 (claude-opus-5-5)
- Setup
- Claude Code subagents, one fresh session per run, plain index.html + styles.css + main.js
- Date
- 2026-11-03
- Without
- No Design Lounge skill or MCP tools and no other design skill; the repository was not read; built in an empty folder.
- With
- The Design Lounge MCP tools as an outside user gets them: start_design first, then recipe, theme, pairing and pieces, and review_page before finishing.
- Blind design rating
- An AI reviewer that had not seen the builds before was shown only screenshots (desktop and phone, first screen and full page), labelled A and B at random, so it could not tell which used the Lounge. It rated fit to the prompt, visual craft, distinctiveness, phone quality and readiness to ship, out of 10 each, 50 in all. It was told the companies are fictional, so a certification or customer claim on a page is invented.
- Library
- The Ledgerly and Fleet builds with the Lounge use the library as of 2026-11-03 (design-lounge-mcp 0.2.6). Crumb & Co is from the first run.
- Review
- review_page from design-lounge-mcp, run locally on every finished build: Chromium at 1280 and 390 wide, axe for accessibility, and the SEO, AEO and GEO checks. Scores start at 100 and lose 14 per fail and 5 per warning.
Does the Design Lounge MCP make better pages?
In a blind design rating, where an AI reviewer saw only screenshots and did not know which build used the Lounge, the build with the Lounge scored higher in 3 of 3 pairs, 118 to 110 out of 150. On the same 3 prompts and the same model, builds with the Lounge MCP had 0 review fails in total against 16 without it, no text below readable contrast against 92 elements, and an accessibility score of 100 on every build.
How many tokens does the Design Lounge MCP use?
About 12.7k tokens for a whole direction: the plan, a recipe, a theme, a type pairing, three piece briefs and one review. Each call is listed above with its size.
The older proof page shows 18 builds of three show-off prompts, with and without the skill. Run it yourself and open an issue with what you get.