QlooVibe

What it does

QlooVibe gives AI coding agents taste, then checks that they used it. Describe your project (or paste a README or design spec) and it returns up to three design directions, each with a named font pair, colour palette, spacing, corner radii, motion, ready CSS and an implementation brief for a coding agent. It works as a web app, a REST API and an MCP server, with no login.

How Qloo is used

A language model proposes a few artists, brands, films, shows or books that the project's audience might love. These starting names are assumptions and are disclosed in the result. Qloo resolves them, and only exact matches are kept. Qloo then supplies ranked decor, ambience and style descriptors, cultural touchstones and affinity scores. QlooVibe filters those descriptors and maps them through a published rule table into design tokens, with the contributing tags and rules shown. All design vocabulary comes from Qloo; no language model writes the tokens. Qloo supplies no font names, colours or CSS, so the mapping is a design hypothesis, and an entity match confirms identity, not that the audience likes it.

After the build: audit and fix

A good brief does not stop an agent shipping generic work. After the agent builds the UI, the audit tool (the audit_ui MCP tool or POST /api/audit) checks the generated code against the chosen direction: colours outside the palette, wrong fonts, starter-template fonts and Tailwind default colours, blue-violet gradients, filler copy and scattered corner radii. It returns specific findings and fixes. These are plain code checks with no model and no Qloo call. A pass means the build follows the direction; it does not prove the design is good.

Measured on a small sample (7 projects, one model, one run): pages written from project text alone passed the audit 0 of 7 times, pages written with the brief 5 of 7, and 7 of 7 after one audit-and-fix round. The brief alone did not remove generic filler copy (2 of 7 pages); the audit-and-fix round did. This shows the loop works on a small sample, not how much it helps in general. Results: docs/audit-benchmark.json in the repo.

Measured

A stored vocabulary benchmark compares Qloo's top-eight decor words with Llama 3.3 70B choosing from the same pooled vocabulary: 12 audiences, three runs. Qloo's sets were more audience-specific (between-audience Jaccard 0.06 vs 0.15), used more distinct words (61 vs 39) and repeated identically (1.00 vs 0.78). This measures vocabulary specificity and repeatability, not design quality.

Built with

Cloudflare Workers and KV, Groq (open model) for suggesting starting audience names with Workers AI as backup, the official MCP SDK, TypeScript and 128 automated tests. Project text is sent to the language model to suggest names; the names are then sent to Qloo. Open source under the MIT licence.

Try it

Live: https://qloovibe.ibnweb3lab.workers.dev (project chooser, three directions, audit tab, hosted MCP endpoint, API docs at /docs). Code: https://github.com/ibnweb3/qloovibe

Built With

• cloudflare-workers
• groq
• mcp
• qloo
• typescript