convez.ai
Skill
plocemourasouza

osforge

plocemourasouza/osforge

Ver no GitHub
InicialPluginScore · de 0 a 100

Production-grade AI development framework: 167 curated skills, 26 agents, rules, hooks, spec commands, OSForge Canvas (local generative UI), and SQLite task board — for Claude Code and Cursor. Next.js + TypeScript + Prisma + Supabase + Bun.

Instalação

$npx convez-ai plocemourasouza/osforge

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
1
Saúde
42%
Último commit
há 1 semana
Versão
v5.3.1
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

🔨 OSForge

Curated agent skills, agents, rules, hooks, commands, and a full AI specialist library for AI-powered development.

OSForge is a production-grade AI development framework with 167 on-demand skills, 26 specialized agents, 13 always-on rules, 9 spec commands, OSForge Canvas (local generative UI for interactive plan/spec review), Python hooks, SQLite local state backend with a cross-project task board, 121 business specialists, and 32 marketing execution workflows — optimized for the Next.js + TypeScript + Prisma + Supabase + Bun stack, with expanded support for mobile, game dev, Rust, Python, and more. Built for Claude Code and Cursor.

"Forging the development environment for AI-powered teams."

📖 Full usage guide → USAGE.md · 💡 10 usage examples → docs/EXAMPLES.md · 🗺️ Architecture map → osforge-architecture.html


Why OSForge?

AI coding agents are only as good as the context they receive. OSForge solves three problems:

  1. Context efficiency — 167 skills in ~12K base tokens (~6% of 200K window). Everything else loads on-demand.
  2. Stack-specific patterns — Core skills tailored for Next.js App Router + Prisma + Supabase + shadcn/ui, with expanded coverage for mobile, game dev, Rust, Python, and cross-platform.
  3. Quality gates built-in — TDD enforcement, security auditing, red team tactics, insecure defaults detection, Reality Check + Quality Control Loop in every agent, and Python hooks at zero token cost.
  4. Local SQLite state — osforge-db CLI persists project state, decisions, blockers, and a task board (waves, dependencies, priorities) with a cross-project board view. Session resumption in ~50 tokens via osforge-db resume. FTS5 full-text search across all decisions cross-project.
  5. OSForge Canvas — a local Thesys-style generative UI: every spec and plan is presented as an interactive browser artifact (cards, tables, dependency graphs, gantt, checklists, approve/edit/reject buttons) instead of terminal markdown. Feedback flows back to the agent as structured JSON. Auto-starts on every Claude Code and Cursor session — zero external APIs.

Quick Start

git clone https://github.com/plocemourasouza/osforge.git
cd osforge
./deploy.sh

The deploy.sh script syncs everything to ~/.claude/ and ~/.cursor/ automatically. See USAGE.md for manual setup and options.


What's Inside

167 Skills (on-demand)

Core & Workflow

#SkillCategory
1TDD Workflow (RED-GREEN-REFACTOR)Core
2Verification Before CompletionCore
3Security Best PracticesSecurity
4Coding Guidelines (Karpathy Rules)Core
5Product-Driven Spec Development (PDD)Core
6Git WorkflowCore
7Clean Code StandardsCore
8Best PracticesCore

Frontend & UI

#SkillCategory
9React & Next.js PerformancePerformance
10Next.js React Expert (9 optimization modules)Performance
11Frontend UI System (shadcn/ui)Frontend
12Frontend Design (colors, typography, animation, UX)Frontend
12bAesthetic Boost (anti-AI-slop, convergence breaking, complexity matching)Frontend
13Tailwind Patterns (v4)Frontend
14UI Design Intelligence (Styles, Palettes, Typography, UX, Charts)Frontend
15Web Design Guidelines (accessibility, UX rules)Frontend
15aTaste Design Dials (DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY dials, GSAP, magnetic, perpetual motion, Bento 2.0)Frontend
15bAesthetic Modes (3 modos: Editorial Minimalist / Industrial Brutalist / Soft Premium)Frontend
15cRedesign Audit (auditoria + upgrades cirúrgicos em projeto existente)Frontend
15dStitch Design Export (DESIGN.md para Google Stitch)Frontend
16Core Web VitalsPerformance
17Accessibility (WCAG 2.1)Frontend
18OpenUI/GenUI LayoutFrontend
19i18n & LocalizationFrontend
20SEOFrontend
21SEO Fundamentals (E-E-A-T, structured data)SEO

Backend & Database

#SkillCategory
22Prisma ExpertStack
23PostgreSQL & Supabase OptimizationPerformance
24Next.js + Supabase AuthStack
25Stripe IntegrationStack
26API Patterns (REST/GraphQL/tRPC + 10 reference modules)Backend
27Database Design (schema, indexing, ORM + 6 reference modules)Backend
28Node.js Best PracticesBackend
29Bun DevelopmentStack
30Server ManagementBackend

Mobile & Game

#SkillCategory
31Mobile Design (touch psychology, platform-specific + references)Mobile
32Game Development (orchestrator + 10 sub-skills)Game
33App Builder (13 project templates)Scaffolding

Security & Compliance

#SkillCategory
34Red Team Tactics (MITRE ATT&CK)Security
35Vulnerability Scanner (OWASP checklists)Security
36Insecure Defaults DetectionSecurity
37Security Threat ModelSecurity
38GDPR / LGPD Data HandlingCompliance

Offensive Security — claude-red (27, on-demand)

Curated/converted subset of SnailSploit/Claude-Red for our web/mobile/cloud AppSec. All passed the skill-metrics audit (median 11.96/12). Authorized testing only.

AreaSkills
Web (16)offensive-sqli, offensive-xss, offensive-ssrf, offensive-idor, offensive-file-upload, offensive-rce, offensive-race-condition, offensive-business-logic, offensive-graphql, offensive-ssti, offensive-xxe, offensive-deserialization, offensive-request-smuggling, offensive-open-redirect, offensive-parameter-pollution, offensive-waf-bypass
Auth (2)offensive-jwt, offensive-oauth
Cloud / AI (2)offensive-cloud, offensive-ai-security
Mobile / Recon (2)offensive-mobile, offensive-osint
Code review / Fuzzing (2)offensive-bug-identification, offensive-fuzzing
Exploit-dev / Utility (3)offensive-toctou, offensive-reporting, offensive-fast-checking

Testing & Quality

#SkillCategory
39E2E Testing (Playwright)Testing
40Testing Patterns (pyramid, unit, integration)Testing
41Web App Testing (Playwright, visual regression)Testing
42Lint and Validate (ESLint, Prettier)Testing
43Code Review ChecklistQuality
44Adversarial ReviewQuality
45Code Review (OSForge Stack)Quality
46Edge Case HunterQuality
47Elicitation EngineQuality
48Readiness GateQuality
49UI Audit (6-pillar Visual Quality Review)Quality
50Differential ReviewQuality

Planning & Ideation

#SkillCategory
51Brainstorming (Socratic questioning)Planning
52Plan Writing (structured planning)Planning
53Spec Builder (Collaborative)Planning
54PRD Builder (Collaborative)Planning
55Architecture Builder (ADR)Planning
56Architecture Patterns (system design + 5 reference modules)Planning
57Epic DecomposerPlanning
58Story ExecutorPlanning
59Phase Discussion (Pre-planning Context Capture)Planning
60Behavioral Modes (agent personas)Planning

Languages & Runtime

#SkillCategory
61Rust Pro (ownership, async, cargo)Language
62Python Patterns (FastAPI, Django, Flask)Language
63Bash Linux (shell scripting)Language
64PowerShell WindowsLanguage

DevOps & Deploy

#SkillCategory
65Deployment Procedures (blue-green, canary, rollback)DevOps
66Vercel DeployDevOps
67Claude CI/CD ActionsDevOps

Documentation & SEO

#SkillCategory
68Documentation Templates (README, API docs, ADR)Docs
69Docs WriterDocs
70Technical Design Doc CreatorDocs
71Doc SanitizationDocs
72GEO / GenAI Optimization (AI search visibility)SEO

Meta & Context

#SkillCategory
73Systematic Debugging (4-phase methodology)Meta
74Performance Profiling (Lighthouse, bundle analysis)Meta
75Predictive Failure AnalysisMeta
76Smart Model DispatchOptimization
77llmfit Advisor (Local LLM Hardware Fit)Optimization
78Context DistillatorContext
79Project Context GeneratorContext

Plus: Doc Shard, Editorial Review, Agent Skills Search, Dispatching Parallel Agents, MCP Builder, Claude API & Agent SDK (TypeScript), Context7 Docs-First, Smart Hooks (Python), AutoRefine Skill, Skill Creator, The Agency (121 AI Specialists), Output Enforcement (anti-truncation/anti-placeholder modifier)

26 Agents (on-demand)

All agents include Reality Check (anti-self-deception) and Quality Control Loop (mandatory self-verification) sections.

Engineering & Code

AgentRole
frontend-engineershadcn/ui + Server Components specialist. Anti-cliché design rules.
backend-engineerPrisma + Supabase + Server Actions
database-architectSchema design, indexing, migrations, ORM selection
mobile-developerReact Native + Flutter, mobile-first design
game-developerGame mechanics, engines, 2D/3D/multiplayer/VR-AR
devops-engineerCI/CD, Docker, infra, pipelines, zero-downtime deploys
performance-optimizerCore Web Vitals, profiling, bundle optimization

Quality & Security

AgentRole
code-reviewerCode quality + YAML-structured review output
code-refactorerRefactoring patterns and clean code
security-auditorTrail of Bits methodology (defensive)
penetration-testerOffensive security, red team, MITRE ATT&CK
test-engineerTesting strategies, TDD, test pyramid design
qa-automation-engineerE2E with Playwright, CI pipelines, visual regression

Planning & Product

AgentRole
orchestratorIntake, triage (21 domains), planning, routing to 26 agents, tracking (always-active meta-agent)
plannerArchitecture, decomposition, story creation (★ Synkra-enhanced)
system-architectSystem design and ADRs
project-plannerDiscovery, task planning, roadmap
product-managerRequirements, user stories, prioritization
product-ownerStrategy, backlog, MVP definition
product-strategy-advisorProduct strategy and roadmap

Investigation & Support

AgentRole
debugger10-step autonomous debugging
explorer-agentCodebase analysis, onboarding, dependency mapping
code-archaeologistLegacy code archaeology, understanding and modernizing old systems
validatorSpec critique + acceptance verification

Documentation & SEO

AgentRole
documentation-writerTechnical docs, manuals, READMEs
seo-specialistSEO, E-E-A-T, structured data, ranking optimization
git-commit-helperConventional commits and release notes

13 Always-On Rules (Cursor)

  • TypeScript Strict Mode — strict: true + noUncheckedIndexedAccess + no any
  • Code Style — Product thinking (PDD), naming conventions, import order
  • Commit Conventions — Conventional commits enforced
  • TDD Enforcement — No production code without failing test first
  • Next.js Patterns — App Router best practices, Server vs Client Components
  • Product Thinking — User-first decisions before technical decisions
  • Security Mindset — Zero-trust by default, fail-safe patterns
  • Agent Skills Reference — How to load and use OSForge skills
  • Orchestrator Awareness — Always check .osforge/status.yaml for work in progress; route complex demands through Orchestrator
  • Artifact Chain — Every planning artifact must have frontmatter with type, status, depends_on; never advance phase without checkpoint approval
  • Intelligent Routing — Silent domain detection + automatic agent + skill selection on every message (21 domains + design taste sub-routing)
  • Anti-AI-Slop — 40 enforced rules preventing generic AI design (typography, color, layout, content, components, performance, meta)
  • Memory Hierarchy — 4-layer CLAUDE.md system (Managed/User/Project/Local) with @include directive, frontmatter paths conditional injection, and explicit override semantics

9 Spec Commands (Claude Code)

Slash commands for spec-driven development (/spec-*):

CommandPurpose
/spec-discoverExplore problem space and gather requirements
/spec-specifyWrite formal specification
/spec-designTechnical design and ADR
/spec-tasksBreak down into implementable tasks
/spec-implementExecute implementation with guardrails
/spec-clarifyClarification loop for ambiguous specs
/spec-checklistPre-ship quality checklist
/spec-constitutionDefine project principles and constraints
/spec-measureDefine and track success metrics

Python Hooks (zero token cost)

HookTriggerPurpose
pre_tool_use.pyBefore any tool callBlocks dangerous commands, protects .env, audit log
post_tool_use.pyAfter Write/EditTypeScript quality gates: console.log, any, @ts-ignore
pre_compact.pyBefore context compactionConversation backup
session_end.pySession endSession logging + macOS notification

🏢 The Agency — 121 AI Specialists (on-demand)

A curated library of 121 AI specialists covering every business and technical function. Each loads only when needed — zero idle context cost.

DivisionAgentsWhen to use
💻 Engineering23Código, arquitetura, segurança, DevOps, SRE, documentação
🎨 Design8UI/UX, sistemas de design, identidade visual, pesquisa com usuários
📢 Marketing26Conteúdo, SEO, redes sociais, growth hacking, ASO + 25 workflows de execução
💰 Paid Media7Google/Meta Ads, PPC, tracking, auditoria de contas + 4 workflows de execução
📊 Product5Roadmap, sprint, pesquisa de mercado, síntese de feedback
🎬 Project Management6Planejamento, Jira/Git workflows, rastreamento de experimentos
💼 Sales8Prospecção, discovery, qualificação, propostas, pipeline + 3 workflows de execução
🛟 Support & Ops6Atendimento, analytics, compliance, resumos executivos
🧪 Testing8QA, API testing, WCAG, performance, validação de entregáveis
🎯 Specialized24Orquestração, compliance SOC2/ISO, blockchain, MCP builder

⚠️ Four agents require mandatory human approval before any autonomous action. See USAGE.md.

32 Marketing Execution Workflows (on-demand)

Integrated from coreyhaines31/marketingskills (MIT). Each workflow provides a step-by-step execution framework that pairs with The Agency's agent personas.

Architecture: Agent (persona: who I am) + Workflow (execution: what I do).

CategoryWorkflowsLocation
CROpage-cro, signup-flow-cro, onboarding-cro, form-cro, popup-cro, paywall-upgrade-croskills/agency/marketing/workflows/
Content & Copycopywriting, copy-editing, content-strategy, email-sequence, lead-magnets, social-contentskills/agency/marketing/workflows/
SEO & Discoveryseo-audit, ai-seo, programmatic-seo, site-architecture, schema-markup, competitor-alternativesskills/agency/marketing/workflows/
Growth & Retentionchurn-prevention, free-tool-strategy, referral-programskills/agency/marketing/workflows/
Strategymarketing-ideas, marketing-psychology, launch-strategy, pricing-strategyskills/agency/marketing/workflows/
Paid Mediapaid-ads, ad-creative, analytics-tracking, ab-test-setupskills/agency/paid-media/workflows/
Salescold-email, sales-enablement, revopsskills/agency/sales/workflows/

All workflows reference .osforge/marketing-context.md as the shared context source. Full agent↔workflow mapping in skills/agency/marketing/workflows/ROUTING.md.


🔬 Agentic AI Patterns (agentic-ai-prompt-research integration)

Integrated from Leonxlnx/agentic-ai-prompt-research (1.6k+ stars). Research-grade patterns extracted from observed Claude Code internal architecture.

4 New Skills

SkillPurposeInspired by Prompt
tool-safety-classifierLLM-powered auto-approval gate for autonomous tool execution. 3 user-customizable sections (allow / soft_deny / environment). Injection defense: compact transcript excludes assistant text.12 — YOLO Classifier
context-compact9-section structured summarization when context > 70%. Three modes (full / partial-recent / partial-up-to). NO_TOOLS_PREAMBLE prevents tool calls during summary. <analysis> scratchpad stripped before delivery.21 — Compact Service
config-critiqueLLM linter for user customizations (new SKILL.md, rules, hooks, agents). Evaluates in 4 axes: Clarity / Completeness / Conflicts / Actionability. Cross-references against baseline.17 — Auto Mode Critique
stuck-recoveryDetect agent loops (same tool failing 3x, no progress in N turns) + save state via osforge-db + propose surgical reset. 4-phase protocol: STOP → SAVE → DIAGNOSE → RECOVER.26 — Stuck Skill (reframed)

1 New Rule + 2 Enrichments

  • rules/memory-hierarchy.mdc (NEW) — Formalizes 4-layer CLAUDE.md system, @include directive (depth 5), frontmatter paths conditional injection, override resolution rules.
  • agents/orchestrator/AGENT.md (enriched) — Added Coordinator Protocol section: synthesize-before-delegate, Continue vs Spawn decision matrix, parallelism rules, real verification philosophy, worker prompt guidance, <task-notification> XML format.
  • claude-code/CLAUDE.md (enriched) — Documented Cacheable Prefix + Cache Boundary + Dynamic Suffix pattern for Anthropic API prompt caching (~10% cost reduction on stable prefix).

Why This Matters

The taste-skill integration improved output quality (anti-slop, premium design). This integration improves agent infrastructure — how agents coordinate, handle context limits, validate customizations, and recover from stuck states. Together they make OSForge production-ready for:

  • Long-running autonomous sessions (auto-mode + context-compact + stuck-recovery)
  • Multi-agent orchestration (Coordinator Protocol + parallelism rules)
  • Enterprise deployments (memory-hierarchy with Managed layer)
  • Open-source contributions (config-critique as PR review gate)

🎨 Design Taste System (taste-skill integration)

Integrated from Leonxlnx/taste-skill (MIT, 10k+ stars). A complete premium frontend engineering toolkit that overrides default LLM design biases via three tunable dials, mode-specific paradigms, and audit-and-upgrade workflows.

The 3-Dial System (taste-design-dials)

Premium frontend at the level of award-winning agency work — tunable on three independent axes:

DialRangeWhat it controls
DESIGN_VARIANCE1 (Predictable Symmetric) → 10 (Artsy Chaos)Layout asymmetry, grid breaking, negative space
MOTION_INTENSITY1 (Static) → 10 (Cinematic GSAP scrolltelling)Animations, spring physics, perpetual loops
VISUAL_DENSITY1 (Art Gallery Airy) → 10 (Cockpit/Packed Data)Spacing, card usage, information density

Default baseline: VARIANCE=8 · MOTION=6 · DENSITY=4. Auto-adjusts to user prompt ("mais minimal" → lower density; "animações fortes" → raise motion).

Includes: Creative Arsenal (50+ premium components), Motion-Engine Bento 2.0 paradigm, double-bezel nested architecture, magnetic micro-physics, liquid glass refraction, GSAP scroll-pinning, AI Tells (forbidden patterns), and the full Pre-Flight Check matrix.

The 3 Aesthetic Modes (aesthetic-modes)

Pick ONE per project and commit. Each mode has its own complete typographic, chromatic, and motion architecture:

ModeVibeUse for
EDITORIAL_MINIMALISTWarm monochrome, document-style, typographic contrastNotion/Linear-style products, knowledge tools, content-heavy SaaS
INDUSTRIAL_BRUTALISTSwiss print or CRT terminal, rigid grids, hazard red accentData-heavy dashboards, telemetry, blueprint UIs, portfolios
SOFT_PREMIUMApple-tier, haptic depth, double-bezel cards, cinematic motionPremium SaaS, AI products, agency sites, Awwwards-tier launches

Audit + Upgrade Loop (redesign-audit)

Different from review-only skills. This skill audits AND applies fixes working with the existing stack — no rewrites. Covers typography, color, layout, interactivity, content, components, iconography, code quality, and the 5 "things AI typically forgets" (legal links, back nav, 404, validation, skip links).

Cross-Cutting Quality Modifiers

  • output-enforcement — anti-laziness directive. Bans // ... rest of code, // TODO, for brevity. Enforces complete output via SCOPE-counting and clean [PAUSED — X of Y complete] breakpoints when token limit looms.
  • stitch-design-export — generates DESIGN.md files compatible with Google Stitch for semantic design system handoff.

How auto-routing works

The rules/intelligent-routing.mdc rule silently detects design intent and bundles the right skills:

"crie uma landing page premium"
  → frontend-design + aesthetic-boost + taste-design-dials + (ask: which mood?)

"estilo Notion / minimalist editorial"
  → aesthetic-modes → EDITORIAL_MINIMALIST

"redesenhar nosso site"
  → redesign-audit → ui-audit (validation loop)

"refatora tudo isso sem placeholder"
  → output-enforcement modifier activates on any agent

Combined with the enhanced anti-ai-slop.mdc rule (40 enforced anti-patterns spanning typography, color, layout, content, components, performance, and meta), every frontend output is automatically gated against generic AI design.


🔍 llmfit Advisor — Local LLM Hardware Fit

Detects your hardware (RAM, CPU, GPU/VRAM) and recommends which local LLMs will actually run well on your machine — with optimal quantization, speed estimates, and fit scoring across 497 models from 133 providers.

Requires: brew install llmfit or cargo install llmfit

Key use cases: LGPD-sensitive data (Essent/Rede Essent Jus), OSystems clients without API budget, and high-volume repetitive dev tasks where local models eliminate API costs.

Source: AlexsJones/llmfit (MIT)


🗄️ Local SQLite State — osforge-db

Persistent project state management via a local SQLite database (~/.osforge/osforge.db). No server, no network, no dependencies beyond Python's built-in sqlite3.

What it stores

TableContent
projectsslug, description, triage, status
phasesper-phase status, skill used, artifact produced
statecurrent phase + resume point (session continuity)
decisionsarchitectural, product, UX, data, security decisions
blockersactive impediments + what they're waiting on
taskstask board: status, wave, depends_on, priority — osforge-db board gives a cross-project view

decisions is backed by an FTS5 virtual table — full-text search across all decisions from all projects in milliseconds.

Core commands

# Session start — load context in ~50 tokens
osforge-db resume my-project

# Record progress
osforge-db set-phase my-project "spec-builder" complete skills/planning/spec-builder docs/specs/feature.md
osforge-db add-decision my-project "Usar Prisma multi-tenant via organizationId" --category=arch

# Session end — mandatory
osforge-db set-resume my-project "Próximo: arch-builder — definir schema de billing"

# Cross-project semantic search
osforge-db search "autenticação OAuth multi-tenant"

# Full project status
osforge-db status my-project

Shell injection in skills

# In any SKILL.md — injects ~50 tokens of precise context
!`osforge-db resume PROJETO_SLUG`

Two scopes

  • Global (~/.osforge/osforge.db) — cross-project state, decision history, no sensitive data
  • Local (.osforge/osforge.db) — per-project, .gitignore-able for sensitive projects (Essent, Rede Essent Jus)

Installed by deploy.sh

deploy.sh installs osforge-db to ~/.local/bin/ and initializes the global database automatically.


osforge/
├── .osforge/              # Project status tracking (status.yaml)
├── claude-code/
│   ├── CLAUDE.md          # Entry point for Claude Code sessions
│   ├── SKILLS.md          # 167 skill triggers (~12K base tokens)
│   └── agents/            # 26 agent definitions (.md)
├── agents/                # 26 agent source files (with Reality Check + Quality Control Loop)
│   ├── orchestrator/      # Meta-agent: AGENT.md + triage-rules + plan-templates
│   ├── frontend-engineer.md  # + Anti-Cliché Design Rules
│   ├── mobile-developer.md   # React Native + Flutter
│   ├── game-developer.md     # Game mechanics, multi-platform
│   ├── penetration-tester.md # Offensive security, red team
│   ├── code-archaeologist.md # Legacy code archaeology
│   ├── explorer-agent.md     # Codebase analysis, onboarding
│   └── ... (26 total)
├── rules/                 # 12 always-on rules (.mdc + .md) for Cursor
│   ├── intelligent-routing.mdc  # 27 domains, design taste sub-routing, automatic skill bundling
│   └── anti-ai-slop.mdc        # 40 enforced rules vs generic AI design
├── commands/              # 9 spec-* slash commands for Claude Code
├── skills/                # On-demand skill library (167 SKILL.md files)
│   ├── agency/            # 121 AI specialists — The Agency (10 divisions) + 32 marketing workflows
│   ├── app-builder/       # 13 project templates (Next.js, FastAPI, Flutter, Electron...)
│   ├── api-patterns/      # REST/GraphQL/tRPC + 10 reference modules
│   ├── game-development/  # Orchestrator + 10 sub-skills (2D, 3D, multiplayer, VR/AR...)
│   ├── mobile-design/     # Touch psychology, platform-specific + references
│   ├── frontend-design/   # Colors, typography, animation, UX + 7 references
│   ├── taste-design-dials/ # 3-dial tunable system (VARIANCE/MOTION/DENSITY) + Creative Arsenal + Bento 2.0
│   ├── aesthetic-modes/   # 3 modos: Editorial Minimalist / Industrial Brutalist / Soft Premium
│   ├── aesthetic-boost/   # Anti-AI-slop compact primer (~500 tokens)
│   ├── redesign-audit/    # Audit + cirurgia em projetos existentes
│   ├── stitch-design-export/ # DESIGN.md generator para Google Stitch
│   ├── output-enforcement/ # Anti-truncation / anti-placeholder modifier
│   ├── ui-design-intelligence/ # Design system spec por indústria (67 estilos, 161 paletas)
│   ├── architecture/      # System design patterns + 5 reference modules
│   ├── database-design/   # Schema, indexing, ORM + 6 reference modules
│   ├── nextjs-react-expert/ # 9 performance optimization modules
│   ├── red-team-tactics/  # MITRE ATT&CK offensive security
│   ├── vulnerability-scanner/ # OWASP checklists
│   ├── rust-pro/          # Ownership, async, cargo
│   ├── python-patterns/   # FastAPI, Django, Flask
│   ├── brainstorming/     # Socratic questioning + dynamic questioning
│   ├── planning/          # 6 planning skills (spec, prd, arch, epics, story, phase)
│   ├── quality/           # 6 quality skills (adversarial, code-review, edge-case, elicitation, readiness, ui-audit)
│   ├── context/           # 4 context skills (distillator, project-context, doc-shard, editorial)
│   └── ... (167 SKILL.md files total)
├── hooks/                 # Python hooks + shell scripts + config
├── mcp/                   # MCP server configs (claude-code.json, cursor.json)
├── scripts/               # Utility scripts
├── docs/                  # Internal documentation
├── osforge-architecture.html  # Interactive visual architecture map
├── deploy.sh              # One-command sync to ~/.claude/ and ~/.cursor/
└── sources/               # Upstream skill collections (gitignored — curation raw material)

Token Budget

ComponentTokensLoaded
SKILLS.md (167 triggers + 13 rules)~12,000Always
Agent definitions (26)~1,500–3,000On invoke
Agency router (SKILL.md)~2,000On invoke
Agency division index (1 of 10)~1,500On invoke
Agency specialist (1 of 121)~2,000–4,000On invoke
Marketing workflow (1 of 32)~1,500–3,500On invoke
Individual SKILL.md files~500–3,000 eachOn invoke
Complex skills with references~3,000–8,000 eachOn invoke
Python hooks0Runtime only
Orchestrator AGENT.md (full routing tables)~5,500On invoke
Planning/Quality/Context skills~500–2,000 eachOn invoke
Base context usage~12,0006% of 200K

🖼️ OSForge Canvas — Local Generative UI

A Thesys-style generative UI, fully local, zero external APIs. The agent writes versioned JSON artifacts; a zero-dependency Bun server (127.0.0.1:4242) renders them as interactive UI in the browser; your feedback flows back as structured JSON the agent reads on its next turn.

Claude/Cursor agent ──writes──▶ ~/.osforge/canvas/artifacts/<id>.json
        ▲                                   │ fs.watch + SSE (live reload)
        │                                   ▼
  reads feedback ◀──writes── Browser viewer (cards, tables, mermaid graphs,
  next turn                  gantt, checklists, forms, approve/edit/reject)
  • Default presentation channel — every spec and plan is emitted as a canvas artifact; the terminal gets a short TL;DR + URL. Opt out with "só texto".
  • Auto-start — SessionStart hooks (Claude Code and Cursor) health-check port 4242 and boot the server if down. Never blocks the session; silent no-op without Bun.
  • 8 block types — heading, markdown, card grids, tables, checklists, forms, decision buttons (with required comments), mermaid (dependency graphs + gantt).
  • Pseudo-streaming — the agent overwrites the artifact as sections complete (revision+1); SSE updates the browser live, no reload.
  • Server: scripts/canvas/server.ts (deployed to ~/.claude/canvas/). Skill: skills/osforge-canvas/ (schema in references/schema.md).

Stack Compatibility

LayerTechnology
FrameworkNext.js 15+ (App Router)
LanguageTypeScript (strict mode)
ORMPrisma
DatabasePostgreSQL via Supabase
AuthSupabase Auth (SSR)
UIshadcn/ui + Tailwind CSS
RuntimeBun
DeploymentVercel
PaymentsStripe
TestingPlaywright + Bun test
AI ToolsClaude Code, Cursor

MCP Servers (Recommended)

// .mcp.json (project) or ~/.claude.json (global)
{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"]
    },
    "supabase": {
      "command": "npx",
      "args": ["-y", "@supabase/mcp-server-supabase@latest", "--read-only"],
      "env": { "SUPABASE_ACCESS_TOKEN": "your-token" }
    },
    "shadcn": {
      "command": "npx",
      "args": ["shadcn@latest", "mcp"]
    }
  }
}

See mcp/claude-code.json and mcp/cursor.json for full configs.


Origins

OSForge was built by evaluating 1100+ agent skills, commands, and patterns across 23 sources, curating e adaptando os mais relevantes em um framework unificado:

SourceRepositoryFocus
Anthropicanthropics/skills ⭐ 63.9kOfficial skills: docx, pdf, pptx, xlsx, mcp-builder, frontend-design
Anthropic (Plugin)anthropics/claude-code/plugins/frontend-design ⭐ 66.5kAnti-convergence aesthetics, aesthetic-boost inspiration, complexity matching (277K+ installs)
Antigravitysickn33/antigravity-awesome-skills ⭐ 7.4k634+ universal skills (curated subset)
Antigravity Kitvudovn/antigravity-kit ⭐ 4.2k32 skills adapted (app-builder, api-patterns, game-dev, mobile-design, red-team, rust, python, etc.) + 15 agents + Reality Check/Quality Control Loop patterns
autoresearchkarpathy/autoresearch ⭐ 39kAutonomous research loop pattern (modify→evaluate→keep/discard→repeat) — inspired autorefine-skill
BMAD-METHODbmad-code-org/BMAD-METHOD ⭐ 35.4kOrchestration patterns: intake, triage, multi-phase planning, quality gates, artifact chains
claude-memthedotmack/claude-mem ⭐ 23.3kPersistent memory for Claude Code
Cloudflarecloudflare/skills ⭐ 233Workers, Agents SDK, Durable Objects
Context Engineeringmuratcankoylan/Agent-Skills-for-Context-Engineering ⭐ 8.1kContext engineering patterns
Curated listsVoltAgent / travisvnAwesome lists used for discovery and curation
Expoexpo/skills ⭐ 878Mobile Expo / React Native
GSD (get-shit-done)gsd-build/get-shit-done ⭐ 34.3kPhase discussion, wave execution, UI audit, STATE.md cross-session memory
llmfitAlexsJones/llmfit ⭐ 6.5kHardware-aware local LLM recommendations
Marketing Skillscoreyhaines31/marketingskills ⭐ 16.8k32 marketing execution workflows: CRO, copywriting, SEO, paid ads, sales enablement, RevOps
Sentrygetsentry/skills ⭐ 173Code review, PR workflow
Supabasesupabase/agent-skills ⭐ 1.1kPostgreSQL optimization, Supabase patterns
Superpowers (obra)obra/superpowers ⭐ 118kbrainstorming, finishing-a-branch, receiving-review, git worktrees, two-stage task review pattern
taste-skill (Leonxlnx)Leonxlnx/taste-skill ⭐ 10kHigh-Agency Frontend — 3-dial tunable system (DESIGN_VARIANCE/MOTION_INTENSITY/VISUAL_DENSITY), GSAP scrolltelling, double-bezel cards, perpetual micro-physics, Bento 2.0 paradigm. Inspired taste-design-dials, aesthetic-modes, redesign-audit, output-enforcement, stitch-design-export.
agentic-ai-prompt-research (Leonxlnx)Leonxlnx/agentic-ai-prompt-research ⭐ 1.6kReverse-engineering research of Claude Code internal architecture. Reconstructed prompts for coordinator, classifier, compaction, memory hierarchy, stuck recovery. Inspired tool-safety-classifier, context-compact, config-critique, stuck-recovery, memory-hierarchy.mdc rule, Coordinator Protocol enrichment, Cache Boundary documentation.
The Agencymsitarzewski/agency-agents ⭐ 31k121 AI specialists across 10 business divisions
Trail of Bitstrailofbits/skills ⭐ 2.4kEnterprise security auditing methodology
UI-UX Pro Maxnextlevelbuilder/ui-ux-pro-max-skill ⭐ 39kDesign intelligence database: 67 styles, 161 palettes, 57 font pairs, 99 UX guidelines, 25 chart types
Vercel Labsvercel-labs/agent-skills ⭐ 19.1kReact, Next.js performance, Core Web Vitals
spec-kit (GitHub)github/spec-kit ⭐ 81.4krequirements-clarify, constitution.md pattern, artifact DAG
OpenSpec (Fission-AI)Fission-AI/OpenSpec ⭐ 23.8kdelta/brownfield spec pattern for incrementally extending existing features
shanraisshan best-practiceshanraisshan/claude-code-best-practice ⭐ 19.7kfrontmatter patterns (model, context:fork, allowed-tools), shell injection, Gotchas sections

Breakdown: 116 SKILL.md files · 26 specialized agents · 121 Agency specialists · 32 marketing execution workflows · 85 trigger entries · 12 always-on rules · 21 detectable domains · 13 project templates · 10 game sub-skills · 9 Next.js optimization modules · 33 GSD commands · ~50 BMAD orchestration patterns · 5 UI-UX Pro Max reference databases · 2 tool patterns (llmfit + autoresearch)


License

MIT — See LICENSE

Author

Paulo Souza — @plocemourasouza

Forging the development environment for AI-powered teams.

Ver o original no GitHub

Bom para

Relacionados