ai-ceo-framework

agent
Security Audit
Pass
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 61 GitHub stars
Code Pass
  • Code scan — Scanned 1 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Open-source multi-agent framework for running a company with Claude Code

README.md

AI-CEO Framework

English | 日本語 | 中文

License: MIT
GitHub stars
GitHub last commit

Run Your Entire Company with Claude Code

15 AI Agents + 5 Production Skills + Approval Pipeline. One-person company, mostly run by automation.


Running in Production

AI-CEO Framework runs a real company. These are the figures we can check, as of 2026-11-11:

Metric Value How it is measured
In production since 2026-02-16 (about 9 months) First commit in the operating repo
Commits made by automation 90% (424 of 470) Last 90 days, commits prefixed auto:
Scheduled jobs 32 Active launchd agents
Departments managed 11 Dev, Marketing, Sales, Finance, Legal, CS, HR, Publishing, Growth, Consulting, BizDev
AI agents shipped here 15 Files in agents/
Reusable skills shipped here 5 Directories in skills/

The running cost is a Claude Code Max plan plus metered usage for headless runs. The metered part is not fixed: it grew as we added automation.

The operating company itself runs 17 agents and 13 skills. The extras are specific to our business (publishing pipelines, outbound sales) and are not distributed here.

An earlier version of this README claimed "98% of business operations", "in production since 2025 (1+ year)" and "~$250/month". The 98% was never measured, and "since 2025" was wrong. They have been replaced with figures that can be checked.

One CEO. Zero employees. Full C-suite powered by Claude Code.


What's Inside

15 AI Agents (agents/)

Each agent has a defined persona, expertise, workflows, output templates, and quality checks.

Agent Role Key Capabilities
CTO Chief Technology Officer Sprint planning, code review, architecture decisions, hotfix management
CMO Chief Marketing Officer Content strategy, SEO, ad campaigns, social media, analytics
CFO Chief Financial Officer Monthly P&L, cost optimization, invoicing, cash flow forecasting
CSO Chief Sales Officer Pipeline management, proposals, lead strategy, CRM
Legal General Counsel Contract review, compliance checks, OSS license audits, ToS/privacy
CS Lead Customer Success Escalation management, FAQ, onboarding optimization, NPS
HR Chief HR Officer Agent skill audits, training plans, performance reviews
Publisher Publishing Director Book planning, writing, quality scoring, multi-channel publishing
Content Engine Content Producer SEO articles, books, LP copy, ad copy, social posts
Growth Growth Hacker Funnel optimization, A/B testing, monetization, pricing
Consulting Consulting VP AI automation consulting, diagnostics, proposals
BizDev Business Development Lead generation, partnership development, upsell strategy
Tax Accountant Tax Advisor Journal entries, tax prep, tax savings, compliance calendar
Morning Digest Daily Briefing Collects all department states, generates CEO morning digest
Setup Wizard Onboarding Interview-based setup, generates all initial config files

5 Skills (skills/)

Reusable, invocable skill definitions that agents call to execute specific tasks.

Skill Purpose
validate-hypothesis 6-phase business hypothesis validation (prevents building things nobody wants)
write-blog Blog article creation with platform-specific scoring (75+ point threshold)
polish-content Content editing and quality improvement for blog posts
upgrade-automation Detect new Claude Code features and upgrade your automation
generate-cover Book cover image generation using HTML+CSS+Playwright

Orchestrator (CLAUDE.md)

The brain of the system. A single CLAUDE.md that:

  • Understands natural language requests and routes them to the right department
  • Manages the approval pipeline for external-facing actions
  • Coordinates cross-department tasks
  • Enforces hypothesis validation gates for new initiatives
  • Maintains a thin context footprint (10-15% utilization)

Steering Files (steering/)

File Purpose
permissions.md Permission levels (read-only / draft / execute), cost thresholds, deploy rules
policies.md Security, quality, cost management, development process, compliance policies

Setup Script (setup.sh)

5-minute setup. Run the script, answer a few questions, and your AI-CEO Framework is live.


Architecture

flowchart TD
    CEO["CEO (You)"]
    ORC["CLAUDE.md\n(Orchestrator)"]
    
    CEO -->|Natural language| ORC
    
    ORC --> CTO["CTO Agent\nSprints, Code Review"]
    ORC --> CMO["CMO Agent\nContent, SEO, Ads"]
    ORC --> CFO["CFO Agent\nP&L, Invoicing"]
    ORC --> CSO["CSO Agent\nSales Pipeline"]
    ORC --> LGL["Legal Agent\nContracts, Compliance"]
    ORC --> PUB["Publisher Agent\nBooks, Multi-channel"]
    ORC --> MORE["+ 9 more agents..."]
    
    ORC --> HYP["Hypothesis\nValidation"]
    ORC --> APQ["Approval\nPipeline"]
    
    HYP -->|Go / No-Go| ORC
    APQ -->|CEO approves| EXE["Execute"]
    APQ -->|CEO rejects| REJ["Reject + Feedback"]
    
    style CEO fill:#f9a825,stroke:#f57f17,color:#000
    style ORC fill:#1565c0,stroke:#0d47a1,color:#fff
    style HYP fill:#7b1fa2,stroke:#4a148c,color:#fff
    style APQ fill:#c62828,stroke:#b71c1c,color:#fff
    style EXE fill:#2e7d32,stroke:#1b5e20,color:#fff

How It Works

The Approval Pipeline

Every external-facing action (emails, social posts, invoices, deployments) goes through:

  1. Agent creates a draft in approval-queue.md
  2. CEO reviews with /ai-ceo:approve <id> or /ai-ceo:reject <id> "reason"
  3. Approved items execute automatically

This prevents AI from sending embarrassing emails or deploying broken code to production.

Hypothesis Validation Gate

Before any new product, ad channel, or significant investment:

  1. Phase 0: Idea origin check (is this customer-driven or self-driven?)
  2. Gate 1: Market existence confirmation (data-backed)
  3. Gate 2: Customer/competitor interviews (3+ companies)
  4. Gate 3: Interview evaluation (fact strength scoring)
  5. Gate 4: Willingness to pay (LOI, pre-orders, written commitment)
  6. Gate 5: Minimum viable test (no-code validation)
  7. Decision: Go / No-Go / Retreat (with learning documentation)

Max 2 retries. If it fails twice, a retreat report is generated. This prevents building things nobody wants.


Quick Start (5 minutes)

1. Clone and setup

# Copy the framework to your project
cp -r ai-ceo-framework-pack/.claude/ your-project/.claude/
cp ai-ceo-framework-pack/CLAUDE.md your-project/CLAUDE.md

# Run the setup script
cd your-project
bash .claude/setup.sh

2. Answer setup questions

The script will ask you:

  • Company name
  • Your name (CEO)
  • Business description
  • Products you're building
  • Tech stack
  • External tools (accounting, CRM, etc.)
  • Which departments to prioritize
  • AI budget

3. Start using it

Open Claude Code in your project directory. Just talk naturally:

> "What's our status today?"
> "Write a blog post about AI automation"
> "Run a dev sprint"
> "Review this contract: [paste]"
> "Generate an invoice for Client X"
> "What are our marketing KPIs?"

Or use explicit commands:

> /ai-ceo:morning          -- Daily briefing
> /ai-ceo:status           -- Quick status check
> /ai-ceo:dev:sprint       -- Run a development sprint
> /ai-ceo:mkt:content-plan -- Monthly content calendar
> /ai-ceo:fin:monthly-report -- Monthly P&L
> /ai-ceo:approve AQ-001   -- Approve a pending item

Directory Structure

After setup, your project will have:

your-project/
  CLAUDE.md                          # Orchestrator (main brain)
  .claude/
    agents/
      cto-agent.md                   # 15 agent definitions
      cmo-agent.md
      ...
    skills/
      validate-hypothesis/SKILL.md         # 5 skill definitions
      write-blog/SKILL.md
      ...
  .company/
    VISION.md                        # Mission & vision
    STATE.md                         # Current business state
    ROADMAP.md                       # Quarterly roadmap
    approval-queue.md                # Pending approvals
    steering/
      permissions.md                 # Permission levels & thresholds
      policies.md                    # Company policies
      brand.md                       # Brand guidelines
      tech-stack.md                  # Tech stack conventions
    products/
      {product-name}/
        STATE.md                     # Per-product state
    departments/
      dev/STATE.md                   # Per-department state
      marketing/STATE.md
      sales/STATE.md
      finance/STATE.md
      cs/STATE.md
      legal/STATE.md
      hr/STATE.md
      publishing/STATE.md
      consulting/STATE.md
    decisions/
      {YYYY-MM}.md                   # CEO decision log

FAQ

Q: Does this work with Claude Code free tier?
A: The framework itself works with any Claude Code plan. However, sub-agents (which use the Agent tool) require Claude Code Max ($100/month) or Claude Code with an Anthropic API key. For optimal performance, Claude Code Max is recommended.

Q: What languages does the framework support?
A: The framework templates are in English. All agent definitions and commands work in English. You can customize agents to work in any language after setup.

Q: Can I add my own agents?
A: Yes. Create a new .md file in .claude/agents/ following the same format (frontmatter + persona + workflows + quality checks). The orchestrator will automatically discover it.

Q: Can I remove agents I don't need?
A: Yes. Simply delete the agent file. The orchestrator gracefully handles missing departments.

Q: How does this differ from just writing a long CLAUDE.md?
A: Three key differences: (1) The orchestrator keeps its context at 10-15% by delegating to sub-agents instead of loading everything into one prompt. (2) The approval pipeline prevents AI from taking unauthorized external actions. (3) Each agent has specialized expertise, quality checks, and output templates that a monolithic CLAUDE.md cannot maintain.

Q: Is my data safe?
A: All data stays in your local .company/ directory. Nothing is sent anywhere except through Claude Code's normal API calls. The framework includes security policies and permission controls out of the box.

Q: Can I use this for a team (not solo)?
A: The framework was designed for one-person companies but works for small teams too. The approval pipeline naturally supports a single decision-maker (CEO). For larger teams, you may want to customize the approval flow.

Q: What if an agent fails?
A: Built-in error handling: retry up to 3 times with feedback, then escalate to the CEO via the approval queue. Error logs are written to each department's error-log.md.


Learn the Background (Japanese)

The operating record behind this framework is written up as a Japanese book: Claude Codeで会社を動かす (free chapters available on Zenn).


Professional Support

(Japan-based companies only)

Need help setting up AI-CEO Framework for your company? JoinClass LLC offers consulting services:

  • Free 30-min AI diagnosis — We analyze your business operations and identify the top 3 areas where AI automation will have the biggest impact
  • Custom setup — We configure the framework for your specific business needs
  • Ongoing support — Monthly advisory for AI operations optimization

Book a free diagnosis →


Before / After

Without AI-CEO With AI-CEO
Setup One giant CLAUDE.md with everything 15 specialized agents + thin orchestrator
Context usage 100% consumed instantly 10-15% orchestrator, delegates the rest
External actions AI sends emails, deploys freely Approval pipeline: draft -> review -> execute
Error handling Errors silently ignored Auto-retry 3x, then escalate to CEO
New initiatives Build first, validate later Hypothesis validation gate (6 phases)
Quality No standards Per-agent quality checks + scoring thresholds
Scaling Rewrite everything Add a new .md file, orchestrator auto-discovers

What This Really Is

This is not a prompt collection. It's a production-grade operating system for running a company with AI. Every agent, every workflow, every quality check was refined through daily use since February 2026. The hypothesis validation skill is the one we rely on most: in August 2026 it was part of the decision to narrow the company to three businesses and shut down five products.

This is the distilled experience of running an entire company with Claude Code -- the failures, the fixes, and the frameworks that survived.


Star History

Star History Chart


Contributing

Contributions are welcome! Whether it's a bug report, feature request, or pull request -- all input is valued.

  • Issues: Found a bug or have an idea? Open an issue.
  • Pull Requests: Fork the repo, make your changes, and submit a PR. Please keep changes focused and include a clear description.
  • Discussions: Questions or ideas that aren't quite issues? Start a discussion.

License

MIT License. See LICENSE for details.


Built with Claude Code. Tested in production. Ready for yours.

Reviews (0)

No results found