Open source Registry · API · MCP · Skills

A registry for taste

Agents can load tools, docs, and skills. They still lack a shared substrate for aesthetic standards—the cultivated judgments that separate living work from generic generation.

API for Taste is an open-source project to publish, discover, and run portable taste packages: moodboards, writing schools, UI dialects, research-backed directions, anti-slop packs—callable over HTTP, MCP, and agent skills.

Registry & surfaces Evals & benchmarks GitHub

Local scaffold today · package format + MCP path in design · heuristic judge, schema-stable for better ones · imperorrp/api-for-taste

Why this exists

Skill packs such as design systems-for-agents (Impeccable and kin) ship one excellent standard for a purpose. That model scales poorly to culture: there are as many serious tastes as there are crafts, curators, labs, and eras. The missing layer is not another monolithic skill—it is a registry of many standards, with a common runtime for agents to search, load, score, retrieve exemplars, and transfer judgment across media.

When generation is cheap, discrimination is scarce. This project treats discrimination as infrastructure.

One skill teaches a house style. A registry lets agents choose—and cite—the court they are in.

Project premise

What you publish

A taste package is a named, versioned, tagged bundle. It may include:

  • 01

    Moodboard & media

    Text, links, images, screenshots, references—and anti-exemplars that define the border.

  • 02

    Constitution

    Short natural-language law: what the standard seeks and refuses, falsifiable in practice.

  • 03

    Context & skills

    Playbooks, checklists, domain notes—skill-like modules bound to the aesthetic, not free-floating prompts.

  • 04

    Tool sequences

    Optional preferred tool-call patterns when enacting this taste (critique → revise → re-score).

  • 05

    Provenance & papers

    Authors, license, citations—optional links to studies and research directions that motivate the package.

Surfaces

HTTP

JSON API & library

Encode, score, transfer today; search and package load as the registry hardens. OpenAPI at /docs.

MCP

Tools for agents

Design target: taste_search, taste_get, taste_score, taste_retrieve, taste_transfer—USB-C for aesthetic standards.

SKILLS

Export as agent skills

Packages compile to skill folders installable beside other agent skills—many tastes, one format.

CLI

Local-first authoring

python -m taste_api for encode, score, transfer, serve—git-friendly packages under packages/.

What agents do with it

Generative UI & product

Load a UI dialect package before scaffolding; score drafts; retrieve anti-exemplars of SaaS cosplay.

Look up by concept

Search tags for “editorial restraint” or “lab figures”; pull exemplars in a curator’s style.

Creative writing

Apply a school or author-adjacent package; score chapters; transfer voice across forms.

Research-backed packs

Packages may cite multimodal aesthetics, preference learning, or anti-slop evals—operations still local to the standard.

Evals, not vibes alone

Taste is profile-relative: the primary question is agreement with a declared package. The project also designs hooks to external suites— creative-writing preference benchmarks, aesthetic perception and bias benchmarks, slop/style report cards—so judges and packages can be compared without pretending universal beauty is settled science. See Evals.

Open source path

Near-term success is adoption: a clear package format, seed packages people actually load, MCP and skill exports that fit existing agent workflows, and evaluation that researchers can extend. Monetization—if any—comes later and must not gate the core.

Collaborators are welcome on packages, judges, adapters, golden sets, and docs. Philosophy and model pages remain for depth; this page is the product front door.

Conceptual model Laboratory Philosophy
Status.
  1. Runtime today: heuristic encode / score / transfer + editorial site + CLI.
  2. Package layout draft under packages/; MCP and skill export are design targets.
  3. Repo: github.com/imperorrp/api-for-taste.