Skip to content

OGPermanerd/relay

Repository files navigation

EverySkill

Protect and grow your IP. Fast.

EverySkill is an AI skills platform built on four architectural layers that turn scattered prompts, workflows, and agent configurations into managed, measurable intellectual property.

The Problem

AI skills live in individual CLAUDE.md files, scattered across laptops, invisible to the organization. When someone leaves, their skills leave too. When someone builds something great, nobody else benefits. When a skill works well, there's no data to prove it. And when the next AI platform arrives, everything starts over.

Four Layers

1. Smart Skills Database

A multi-tenant, privacy-scoped skills repository that goes beyond storage. Skills are discovered through semantic search (not just keywords), and the system analyzes your actual work patterns to recommend the skills that will have the most impact on what you're doing right now.

  • Multi-tenant isolation with row-level security and subdomain routing
  • 4-tier visibility: personal, team, tenant-wide, or public
  • Semantic search via Ollama embeddings + pgvector cosine similarity
  • Work-activity analysis (Gmail patterns today, browser history and more channels coming) matched against the skill catalog
  • Duplicate detection on upload to prevent redundant skills
  • Quality-gated publishing with AI review and admin approval workflows

2. IP Stewardship & High Velocity Growth

The sum total of a tenant's (or user's) skills is intellectual property. This layer protects it, measures it, and makes it better through continuous feedback loops.

  • Usage tracking with per-skill metrics (total uses, unique users, hours saved)
  • Quality scoring via automated AI review (clarity, completeness, quality)
  • Feedback sentiment from thumbs up/down with aggregated trends
  • Training data — authors seed golden input/output examples; real usage captures more (with dual consent)
  • Token/cost measurement — actual cost per skill execution across models
  • Benchmarking — cross-model quality comparison with blinded AI judge
  • Suggestion-to-fork pipeline — user suggestions auto-generate improved versions
  • IP protection — companies retain institutional knowledge when employees leave; employees retain personal skills when they move on

3. AI Independence

Skills are portable text, training data is model-agnostic input/output pairs, and benchmarking already compares models head-to-head. The architecture ensures no lock-in to any single AI provider.

  • Model-agnostic skill format — markdown-based, works with any LLM
  • Cross-model benchmarking — evaluate skills across Sonnet, Haiku, and any future model
  • Model-agnostic training data — golden examples are input/output pairs, not fine-tuning artifacts
  • Roadmap: multi-platform execution (OpenAI, Gemini, Llama) with platform-agnostic skill translation

4. Universally Integrated Access

Low friction equals adoption. Skills are accessible wherever you work — browser, prompt, code, or API — with zero context switching.

  • Web application — full CRUD, feedback, suggestions, training, benchmarking, admin controls
  • In-prompt (MCP) — search, execute, track, and give feedback without leaving Claude
  • In-code (hooks) — PostToolUse hooks for automatic tracking, feedback prompting, and training data capture
  • REST API — programmatic access for tracking, feedback, health checks, and integrations
  • Roadmap: cross-AI-platform access (same skill from ChatGPT, Gemini, Claude)

Tech Stack

Layer Technology
Framework Next.js 16 with React 19
Language TypeScript (strict mode), ~50k LOC
Database PostgreSQL 16 + pgvector + Drizzle ORM
Auth Auth.js v5 (Google Workspace SSO)
AI Anthropic SDK, Ollama (nomic-embed-text)
MCP @modelcontextprotocol/sdk + mcp-handler
Monorepo Turborepo + pnpm workspaces
Styling Tailwind CSS v4
Charts Recharts
Testing Playwright (E2E), vitest (unit)
Deployment PM2, Caddy, LXC on Hetzner VPS

Monorepo Structure

everyskill/
├── apps/
│   ├── web/                # Next.js web application
│   └── mcp/                # MCP server (stdio transport)
├── packages/
│   ├── core/               # Shared types and constants
│   ├── db/                 # Drizzle schema, migrations, services
│   ├── storage/            # File storage abstractions
│   └── ui/                 # Shared React components
├── docs/                   # Architecture, infrastructure, guides
│   ├── ARCHITECTURE.md     # 4-layer architecture deep dive
│   ├── INFRASTRUCTURE.md   # Deployment and environments
│   ├── CONTRIBUTING.md     # Developer guide
│   └── API.md              # API reference
└── .planning/              # Roadmap, phase plans, state tracking

Quick Start

Prerequisites

  • Node.js 22+
  • pnpm 9+
  • PostgreSQL 16+ with pgvector extension

Setup

git clone <repository-url>
cd everyskill
pnpm install
cp apps/web/.env.example apps/web/.env.local  # configure DATABASE_URL, auth credentials
pnpm db:migrate                                # run all migrations
pnpm dev                                       # start dev server on :2002

Key Commands

Command Description
pnpm dev Start development server
pnpm build Build all packages
pnpm db:migrate Run database migrations
cd apps/web && npx playwright test Run E2E tests
./deploy.sh staging Deploy to staging
./deploy.sh promote Promote staging to production

Environments

Environment URL Port
Production everyskill.ai 2000
Staging staging.everyskill.ai 2001
Development localhost:2002 2002

Documentation

Milestones

Version Shipped Highlights
v1.0 2026-01-31 MVP: skill CRUD, MCP integration, search, ratings
v1.1 2026-02-01 Quality scorecards, E2E test coverage
v1.2 2026-02-02 Two-panel UI redesign, keyboard navigation
v1.3 2026-02-04 AI review, semantic similarity, forking, cross-platform install
v1.4 2026-02-06 Employee analytics, remote MCP, API keys
v1.5 2026-02-08 Production deployment, multi-tenancy, RBAC, notifications
v2.0 2026-02-08 Quality-gated publishing, conversational MCP, drift detection
v3.0 2026-02-13 AI discovery, hybrid search, homepage redesign, preferences
v4.0 2026-02-14 Gmail workflow diagnostic, work-activity skill recommendations
v5.0 2026-02-15 Feedback loops, training data, benchmarking, cost measurement

License

Proprietary. Internal use only.

About

No description, website, or topics provided.

Resources

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages