founder-os

Run your company with an AI team. No servers, no SaaS dashboards, no complex tooling — just markdown files and Claude Code.

What You Get (All From Markdown Files)

Clone this repo, answer a few questions, and you immediately have:

An entire AI team with specialized roles. Not one generic AI — a software architect, product manager, engineer, GTM director, content strategist, technical writer, sales advisor, and more. Each agent has deep context about your company, your tech stack, your market, and your codebase.

A work queue you can fire and forget. Queue up tasks, walk away, come back to completed work with detailed reports. The system picks the highest-priority task, assigns it to the right specialist, and executes autonomously.

A research-before-code pipeline. Unsure about an approach? Queue a research task. The AI investigates, writes a plan document, and waits for your approval before a single line of code is written. You can send it back for revisions. When you approve, an implementation task is auto-created.

Agents that verify their own work. Every task goes through goal-backward verification — the agent re-reads acceptance criteria, traces through the code, and confirms everything actually works before marking complete. No more "it compiles but doesn't do the thing."

A natural language interface. You don't need to memorize commands. Just say "what should I do next?" or "I need to build user auth" and the system figures out the right action. The slash commands are power-user shortcuts, not prerequisites.

Institutional memory that compounds. Learnings accumulate over time. Patterns get discovered. Mistakes get recorded. Your AI team gets smarter about YOUR codebase with every task it completes.

Full founder control. Nothing ships without your review. You approve plans, review completed work, and make the strategic calls. The AI handles execution — you handle direction.

The Superpowers

Here's what's wild — you get all of this with zero infrastructure. No CI/CD pipelines, no project management SaaS, no API integrations. It's all markdown files that Claude Code reads and follows.

Superpower	What it means	How it works
Autonomous execution	Queue work, walk away, come back to results	Work queue with priority-based pickup
Specialist agents	Domain experts, not one generic AI	Agent templates with role-specific context
Research pipeline	Think before you code	Research → plan → approve → implement
Self-verification	Agents check their own work	Goal-backward verification on every task
Dependency management	Work flows in the right order	Tasks can block other tasks
Context injection	Every agent knows your company	Company config, tech stack, learnings loaded automatically
Quality sizing	Tasks scoped to prevent quality degradation	Single context window, 3-5 files max
Completion reports	Know exactly what was done	Detailed markdown reports for every task
Natural language	Just talk to it	Claude interprets intent, picks the right command
Institutional memory	Your team gets smarter over time	Learnings persist and inform future work
Extensible	Build your own commands	Command creation wizard, composable primitives
Zero infrastructure	No servers, no SaaS, no setup	Everything is markdown files in your repo

Quick Start

1. Clone the framework

git clone https://github.com/cloudrepo-io/founder-os.git my-company
cd my-company

2. Run the setup wizard

# In Claude Code
/init

Answer a few questions about your company — name, tech stack, which team members you want, market context — and the framework configures itself.

3. Start working

# Queue up some work
/queue:add Build user authentication

# Or investigate something first
/queue:research Evaluate authentication approaches for our stack

# Let the AI team handle it
/queue:work

# Check your dashboard
/founder:dashboard

Or just talk naturally:

"What should I work on next?"
"I need to build a payment integration"
"How's the project going?"
"Review what the team did today"

Commands

Command	What it does
/queue:add	Add an implementation task
/queue:research	Add a research/investigation task
/queue:work	Pick up and complete the next task
/queue:status	View current queue state
/queue:approve	Approve a research plan
/queue:revise	Request changes to a plan
/team:add	Add a new AI team member
/team:list	List your current team
/founder:dashboard	Overview of company state
/founder:review	Review completed work
/config:set	Update company config
/learnings:add	Capture institutional knowledge
/create-command	Create a new custom slash command

Available Agents

Agent	Role
staff-software-architect	System design, architecture decisions, technical leadership
staff-product-manager	Product strategy, requirements, prioritization
staff-engineer	Implementation across any language/framework
staff-gtm-director	Go-to-market strategy, marketing coordination
staff-content-strategist	Content creation, copywriting, SEO
staff-technical-writer	Documentation, tutorials, API reference
staff-sales-advisor	Sales strategy, pipeline, founder coaching
staff-api-tester	API testing, test automation
staff-security-engineer	Security testing, vulnerability discovery

How It Works

Two Paths to Getting Work Done

Known work — you know what needs building:

/queue:add → /queue:work → done → /founder:review

Needs investigation — scope unclear, multiple approaches:

/queue:research → /queue:work → plan → /queue:approve → /queue:work → done

The research pipeline ensures major changes get proper thought before code is written. Plans go through a review loop until you're satisfied.

What Happens Under the Hood

1. You queue a task (or just describe what you need)
2. The system loads your company context, tech stack, and accumulated learnings
3. The right specialist agent picks up the task
4. The agent executes, verifying against acceptance criteria as it goes
5. On completion, a detailed report is generated
6. You review and approve (or request changes)

All of this is driven by markdown files that define the workflow. There is no backend. There is no database. The queue is JSON files. The agents are markdown personas. The commands are markdown instructions. And it all just works because Claude Code reads and follows them.

Directory Structure

your-company/
├── .claude/
│   ├── commands/           # Slash commands
│   │   ├── queue/          # Queue management (6 commands)
│   │   ├── team/           # Team management
│   │   ├── founder/        # Founder tools
│   │   ├── config/         # Configuration
│   │   ├── learnings/      # Knowledge capture
│   │   ├── create-command.md  # Command creation wizard
│   │   └── init.md         # Setup wizard
│   ├── agents/             # Your AI team
│   │   └── _templates/     # Agent templates
│   ├── queue/              # Work queue system
│   │   ├── active.json     # In-progress work
│   │   ├── backlog.json    # Pending work
│   │   ├── completed/      # Archived work
│   │   ├── plans/          # Research plan documents
│   │   └── reports/        # Completion reports
│   ├── learnings/          # Institutional memory
│   ├── team.json           # Team roster
│   ├── company.yaml        # Company configuration
│   └── founder-actions.md  # Human task tracker
├── CLAUDE.md               # AI team operating manual (the brain)
└── README.md               # This file

Philosophy

Just Markdown Files

This is the core insight: Claude Code already knows how to read files, follow instructions, and execute work. You don't need a framework with dependencies, a build step, or an API. You just need well-structured markdown files that tell it what to do and how to do it.

Every "feature" in founder-os is a markdown file:

• Agent with deep domain expertise? A markdown file with a persona and context.
• Autonomous work queue? A markdown file describing the pickup-execute-complete flow.
• Research pipeline with approval gates? Markdown files defining each step.
• Quality verification? A checklist in the CLAUDE.md that agents follow.

When you want to add a new capability, you create a new markdown file. That's it.

Founder + AI Partnership

You're not replacing yourself — you're amplifying yourself. The AI team handles implementation, research, content, documentation, and testing. You handle strategy, decisions, and reviews. The queue system is the interface between you and your team.

Institutional Memory

Your AI team gets smarter over time. Learnings accumulate, patterns get discovered, and mistakes get recorded. A fresh Claude session isn't starting from zero — it reads your learnings and benefits from everything that came before.

Contributing

This is an open-source framework. Contributions welcome:

• New agent templates
• Additional commands
• Documentation improvements
• Example configurations

License

MIT — see LICENSE for details.

---

Built by CloudRepo

This framework powers how we build new features at CloudRepo — a universal artifact repository for all your packages, containers, and dependencies. If you're building libraries, tools, or apps that you want to share with the world, come check it out.