Multi-Agent Plugins

Per-stack AI engineering-lifecycle toolkits (Claude Code plugin marketplace) for the multi-agent pipeline - iOS, Android, frontend (React), and backend, each project-configurable. Enable your stack's plugin per repo.

A Claude Code plugin marketplace of per-stack engineering toolkits. Each plugin answers one question for one stack: how do you create, evolve, fix, review, build/test and ship code here, without hardcoding any single project's names, paths or tokens.

  • 5 plugins, 238 skills: iOS/SwiftUI (143), Backend (32), Android/Compose (29), Frontend/React (24), Common (10)
  • Every plugin has the same anatomy: an index router, reference skills (how this stack is built), workflow skills (what to do, step by step), tools (Figma, Code Connect, docs) and knowledge (vendor/API depth)
  • Component and Figma-to-code work lives here - the pipeline dispatches into these plugins instead of bundling component skills
  • Authored once in the pipeline repo; the marketplace is a generated, versioned publish artifact whose patch version bumps only when a skill set actually changes
  • Project-scoped enablement: a repo enables its stack in .claude/settings.json, and the pipeline detects the stack and uses whatever is enabled
ArchitecturePrivate Repository
6
Plugins
252
Skills
iOS 0.10.0 · Common 0.2.8 · Backend 0.1.5
Versions
iOS · Android · React · Backend · Analyst
Stacks
WIP
Status

Tech Stack

Claude CodePlugin MarketplaceSwiftUIJetpack ComposeReactTypeScriptFigma

Features

Four-layer skill anatomy

Every plugin splits knowledge from procedure, so the model loads only what a task needs.

  • index - a router: reads the request, names the skills that apply, and stops
  • reference (descriptive) - how this stack is built: architecture, component contract, state, navigation, forms, styling, file placement. Read before writing code
  • workflow (imperative) - numbered procedures with gates: create, evolve, fix, review, branch-and-pr. Read while doing the work
  • knowledge - vendor and API depth, generated from the pipeline's shared skill tree
  • Nothing project-specific is hardcoded: scheme names, token namespaces, module paths and test runners come from config

ai-ios-toolkit - 147 skills

The flagship: a full SwiftUI engineering lifecycle, native-first.

  • Reference (15): architecture, component (Configuration/View/+Modifiers), data-model (entities, request/response models, mappers), networking (generated OpenAPI client), navigation, overlays, bottom-sheets, forms, state (@Observable), styling/design tokens, previews, file-placement, code-style, build-and-test
  • Workflow (23): create-component (Figma → SwiftUI, 1:1 with the mapped Code Connect component), create-screen (Entity → Repository → Mapper → UseCase → ViewModel → Scene), evolve-component, fix-bug, branch-and-pr, backlog, the figma-* family and the performance-* tour
  • Tools (5): figma-utility, figma-setup, code-connect, component-docs, component-wiki
  • Knowledge (98): SwiftUI, Swift concurrency, SwiftData, Swift Testing, HIG, App Store review, WidgetKit, Live Activities, StoreKit, security, Xcode build analysis

ai-android-toolkit - 30 skills

The Compose counterpart, same anatomy and the same workflow names.

  • Reference (8): component, state, navigation, forms, styling, overlays, bottom-sheets, build-and-test, written for Compose (MutableStateFlow + collectAsStateWithLifecycle, type-safe routes, Material 3 tokens)
  • Workflow (4): create-component, evolve-component, fix-bug, branch-and-pr
  • Tools (3): figma-utility, code-connect (Compose), component-docs
  • Knowledge (13): Kotlin coroutines, Compose components/navigation/testing, Room, Retrofit, Gradle KTS, Play Store compliance, performance, security
  • Deliberately name-identical to the iOS toolkit: the pipeline's phases stay platform-agnostic and only the enabled plugin changes

ai-frontend-toolkit - 25 skills

React/TSX engineering lifecycle: create, evolve, fix, review, ship.

  • Reference (6): component architecture, state (hooks + Query), forms (React Hook Form + Zod), styling (design tokens), routing (Next.js / React Router), build-and-test (Vitest + Testing Library + Storybook)
  • Workflow (4): create-component (Figma → React), evolve-component (drift-heal + extend, human-gated), fix-bug, branch-and-pr
  • Tools (3): figma-utility, code-connect (React), component-docs (Storybook)
  • Knowledge (10): Next.js App Router, modern CSS, Tailwind, semantic HTML, web performance, web accessibility, web testing, TypeScript patterns

ai-backend-toolkit - 33 skills

Service-side lifecycle with the endpoint as the unit of work.

  • Reference (6): api (REST/GraphQL/tRPC selection, response shape, versioning, pagination), auth, data, validation, security, testing
  • Workflow (4): create-endpoint, evolve-endpoint, fix-bug, branch-and-pr
  • Tools (2): openapi-utility (spec as source of truth), api-docs
  • Knowledge (21): FastAPI, Node backend patterns, database patterns, REST API design, API security, Docker, CI/CD, observability, Python patterns
  • Pairs with the mobile toolkits: the same OpenAPI spec that generates the iOS service client is the one this toolkit evolves

ai-common-toolkit - 11 skills

The genuinely stack-agnostic set, enabled alongside a stack plugin so each stack toolkit stays lean.

  • council - structured disagreement for go/no-go calls and ambiguous tradeoffs
  • search-first - check what already exists before writing a custom implementation
  • backlog - a registry for deferred-but-approved work, so \"later\" never means lost
  • agent-introspection-debugging - capture, diagnose, contained recovery when an agent loops or keeps failing
  • humanizer - strips AI writing tells from PR bodies, commits and docs
  • skill-creator - the house rules for authoring skills
  • accessibility audit and Firebase, used across mobile, web and backend

ai-analyst-toolkit — 5 skills

Evidence gathering for a decision that needs sources rather than opinion: GitHub issue and release archaeology, a registry of what was already established, and community signal. Small on purpose - it grows only when a question keeps coming back without an answer.

Generated, not hand-maintained

The marketplace is a build output, which is why the plugin set never drifts from the pipeline.

  • Vendor/knowledge skills are authored once in the pipeline repo and consumed by the pipeline's own phases
  • The generator rebuilds each plugin's knowledge/, regenerates plugin.json, and bumps a patch version only when that plugin's skill set actually changed
  • Cross-stack skills route to the common toolkit; Apple/Xcode-only skills route to iOS alone
  • Authoring directly in a plugin's knowledge/ is a mistake by construction: the next rebuild overwrites it
  • The pipeline's sync command runs this rebuild and pushes the marketplace in the same pass

Per-project switching

Enable a stack's plugin in the repo's .claude/settings.json.

  • No runtime toggle needed; the multi-agent pipeline detects the stack and uses whatever is enabled
  • Consumers pick up new plugin versions through the pipeline's update command, which refreshes the marketplace
  • Collaborators aligned with multi-agent-pipeline / multi-agent-toolkit-mcp so access is bundled, not requested per-repo