onemanagency
v4.1.1
Published
Sistema de execução de projetos via IA. OneManAgency - funciona em qualquer IDE e adapta-se a qualquer projeto.
Maintainers
Readme
OneManAgency (v4.1)
Open-source framework to build software like you have a full agency working for you.
Multi-LLM Orchestration, native Context Engineering, PIV Loop, and TDD for Claude Code, OpenCode, Cursor, Windsurf, Aider, Hermes Agent, Roo Code, and Gemini CLI.
"Don't write prompts. Architect systems."
Why I Built This
I build SaaS solo. End to end.
Every time I used Claude Code, Cursor, or any LLM to build a real project, the same thing happened:
- Context window filled up. AI forgot the architecture.
- AI started writing generic purple-gradient design (the "AI Slop").
- Code shipped without tests. Broke in production.
- I asked for one feature. AI rewrote the entire project.
Enterprise tools were too heavy. So I took the best methodologies from Silicon Valley (Y Combinator, Spec-Driven Development, TDD) and built OneManAgency (OMA).
The core idea: The complexity lives in the system, not in your workflow. You keep your IDE. Your AI follows corporate rules invisibly.
How OMA Fixes AI
PIV Loop (Plan-Implement-Validate)
The problem: AI plans and codes in the same breath. Context window fills up. AI forgets architecture. Generates "AI Slop."
The fix: OMA forces a hard stop between planning and coding. Agent plans, generates HANDOFF.md, and warns: "Clear the chat." You clear. AI returns with empty context, laser-focused on execution.
Result: Zero hallucination. Zero rewrites.
3-Gate Review Triad
The problem: You ask AI for a feature. It builds everything. Half of it is useless. You waste tokens, time, and money.
The fix: Before code touches your repo, every feature passes through 3 filters:
- CEO — "Does this generate revenue on day 1?" Cuts scope.
- Eng — "Does the schema hold? Any edge cases?" Locks architecture.
- Design — "Is this AI Slop?" Forbids generic design.
Result: Features that survive the Triad are worth building.
Architecture Overview
Quick Start
- Install globally on your machine (terminal):
npm install -g onemanagency@latest- Create a new folder and open it in your AI-powered IDE (Claude Code, Cursor, Roo Code, etc.):
mkdir my-new-saas && cd my-new-saas- Initialize OMA by typing the slash command directly in your IDE's AI chat:
/oma-initThe installer auto-detects your IDE (.claude, .opencode, .cursor, .roo, .gemini, .windsurf, .aider, .cline) and injects skills there.
17 Specialized Personas
OMA uses strict Agent Definition Files in src/agents/. These are files that tell exactly what the agent hates and how it operates. 17 personas triggered by PIPELINE.md via metadata Agent: <name>:
| Category | Agents | |---|---| | Implementation (5) | Frontend Specialist · Backend Specialist · Database Architect · DevOps Engineer · Mobile Specialist | | Design & Content (2) | Design Specialist · Copywriter Specialist | | Quality (5) | Code Reviewer · Accessibility Auditor · Performance Engineer · Reality Checker · Test Engineer | | Specialists (5) | Security Auditor · SEO Specialist · MCP Builder · Chatbot Specialist · Lead Orchestrator |
9 Tested Playbooks
The pipeline-generator has ready playbooks for each project type:
| Playbook | Type | Stack | |---|---|---| | A | SaaS Full-Stack | Next.js + Supabase + Vercel | | B | Static Landing Page | HTML/CSS + Vercel/Netlify | | C | Landing Page Next.js | Next.js + Vercel | | D | Python Automation | Python + FastAPI + Cloud Run | | E | Low-Ticket / Infoproduct | HTML + Kiwify Checkout | | F | Data Pipeline / ETL | Python + Airflow/Prefect + BigQuery | | G | Mobile React Native | React Native + Expo + EAS | | H | WhatsApp Chatbot | Node.js + Baileys/Botpress | | I | Hybrid / Monorepo | Turborepo + shared packages |
How It Works in Practice
1. Setup
Crie uma pasta vazia e abra ela na sua IDE com IA ligada. Chame o Engenheiro de Infraestrutura pelo chat:
/oma-init2. Socratic Interview
OMA doesn't give you a passive form. The agent assumes the persona of a Y Combinator Partner. If you say "I want an app with 50 features", the AI asks: "What's the real pain? Let's focus only on the feature that generates revenue on day 1."
3. The Barrier (Triad)
The briefing passes through 3 automatic filters:
- CEO Review → Generates
PRD.md - Eng Review → Defines schema and data flow (
ARCHITECTURE.md) - Design Review → Defines tokens, typography, and anti-generic motion rules (
UI-SPEC.md)
4. Execution (PIV Loop)
The Pipeline Generator slices everything into atomic tasks. You invoke oma-executor.
The AI plans the task, writes to disk, and asks you to clear the screen. You clear. On return, the Frontend agent (focused on accessibility and Tailwind) or Backend agent (focused on TDD) takes action. No hallucinations.
Key Features
| Feature | Description |
|---|---|
| Context Engineering | Project state persisted on disk (HANDOFF.md, STATE.md), not in chat RAM |
| PIV Loop | Plan → /clear → Implement → Validate. Prevents hallucination. |
| 3-Gate Review | CEO + Eng + Design cut useless features before code is written |
| TDD Iron Law | Backend agent is forbidden to write production code without a failing test first |
| Batteries Included | MCPs auto-configured: Playwright, Context7, Sequential Thinking, Memory, Fetch |
| Offline-First | Works without internet after install. Native FileSystem, not shell scripts. |
| Cross-IDE | Claude Code, Cursor, OpenCode, Windsurf, Aider, Gemini CLI, and more |
| Progressive Disclosure | Skills load only essential (~200 lines), playbooks load on demand |
Bring Your Own Agents
OMA triggers agents via metadata Agent: <name> in PIPELINE.md. Any .md file with frontmatter name: in .agents/agents/ is eligible. You're not limited to the 17 core personas:
| Library | Stars | When it helps | How to integrate |
|---|---|---|---|
| agency-agents | 73-94k | 144 agents in 12 divisions | Copy .md to .agents/agents/ |
| GStack (Garry Tan, YC) | — | 23 personas focused on product validation | Copy agent to .agents/agents/ |
| Superpowers | — | 15 composite skills (TDD, debugging) | Use as skill in .agents/skills/ |
Inspiration (Prior Art)
I didn't invent the wheel. OMA is a synthesis of the brightest minds in Agentic Engineering and Design:
Architecture & Product Management
- GStack (Garry Tan) — Inspired our Review Triad
- Get-Shit-Done — Inspired our lightweight file persistence
- Spec-Kit — Validated Spec-Driven Development
Engineering
- Superpowers — Base of our Backend Specialist (TDD Iron Law)
- Agency-Agents — Taught us that "shallow roleplay" doesn't work
Anti-"AI Slop"
- Impeccable & Taste-Skill — Defense against AI Slop
- Emil Kowalski's Philosophy — Backbone of our Design Specialist
- Huashu Design — Rapid prototyping insights
If you like OMA, please consider giving a Star to the repos of these giants. We stand on their shoulders.
Continuous Integration (24/7 Coworking)
Since OMA uses files on disk (Context Engineering) instead of chat memory, it's the perfect engine for autonomous terminal orchestrators:
- AionUi — Instantiate multiple terminals side by side. AionUi is the "office", OMA is the "method".
- Hermes Agent / OpenClaw — Run the agency on a VPS and command agents via Telegram or Discord.
"You don't level up your code by asking nicely. You level up by building systems that forbid bad code."
MIT License | Made by a builder, for builders.
