Run an autonomous AI company from your terminal.
Live Demo · Website · Watch Video
What is auto-co?
A bash loop that calls Claude Code every 2 minutes. 14 AI agents (CEO, CTO, Engineer, Designer, QA, Marketing...) debate, decide, build, and deploy software — 24/7, without you.
It's not a chatbot. It's not a framework. It's ~50 lines of bash that turn Claude Code into a self-running company.
read consensus → pick agents → execute → update consensus → sleep → repeat
State lives in markdown files. Everything survives restarts. The only dependency is Claude Code.
Quick Start
Prerequisites: Claude Code installed and working.
npx create-auto-co my-company
cd my-company
make startWatch it work: make monitor
That's it. First cycle starts in ~30 seconds. Cost: ~$1.80/cycle.
From a template
npx create-auto-co my-saas --template saas
Templates: saas (Next.js + Supabase + Stripe), docs-site (MDX + Vercel), api-backend (Express + Railway).
Clone directly
git clone https://github.com/NikitaDmitrieff/auto-co-meta cd auto-co-meta cp .env.example .env # optionally configure make start
What to Expect
| Cycles | What happens |
|---|---|
| 1-3 | CEO assembles team, evaluates product ideas, picks a direction |
| 4-8 | Architecture decisions, first code, initial deploy |
| 9-20 | Iterating: features, fixes, landing page, pricing |
| 20+ | Distribution, user acquisition, self-improvement |
Human input needed: ~once every 20-30 cycles (for credentials, spending decisions, or legal questions). Escalation requests come via Telegram if you set it up.
The Agents
14 agents, each modeled on a real-world expert. Each cycle picks 3-5 relevant ones.
| Layer | Agent | Modeled after | Role |
|---|---|---|---|
| Strategy | CEO | Jeff Bezos | Priorities and direction |
| CTO | Werner Vogels | Architecture and tech choices | |
| Critic | Charlie Munger | Veto bad ideas | |
| Product | Product | Don Norman | UX and features |
| Design | Matias Duarte | Visual design | |
| Interaction | Alan Cooper | User flows | |
| Engineering | Fullstack | DHH | Write and ship code |
| QA | James Bach | Test strategy | |
| DevOps | Kelsey Hightower | Deploy and infra | |
| Business | Marketing | Seth Godin | Positioning and distribution |
| Operations | Paul Graham | User acquisition | |
| Sales | Aaron Ross | Pricing and conversion | |
| CFO | Patrick Campbell | Financial model | |
| Intelligence | Research | Ben Thompson | Market research |
Agent definitions live in .claude/agents/. Edit them to change behavior.
Configuration
Copy .env.example to .env. The only required setting:
# .env ANTHROPIC_API_KEY=sk-ant-... # if not using Claude Code's built-in auth MODEL=opus # or sonnet for cheaper cycles (~$0.50)
Everything else has sensible defaults. See .env.example for advanced options (loop interval, idle detection, Telegram notifications, webhooks).
How It Works
auto-loop.sh
├── reads memories/consensus.md ← shared state ("relay baton")
├── builds prompt with PROMPT.md ← instructions for Claude
├── calls `claude -p` with prompt ← Claude Code does the work
├── agents update consensus.md ← next action for next cycle
├── appends to state/*.jsonl ← structured logs (decisions, tasks, artifacts)
└── sleeps → repeats
Key files:
auto-loop.sh— the entire loop (~3000 lines with monitoring, error handling, adaptive frequency)PROMPT.md— system prompt sent to Claude each cyclememories/consensus.md— the "relay baton" that carries state between cycles.claude/agents/*.md— agent persona definitionsMakefile— all commands (make start,make monitor,make status, etc.)
No database. No server. No framework. Just files, git, and a bash loop.
Built With auto-co
These products were built entirely by auto-co instances, from idea to deployment:
| Product | What it does | Live at |
|---|---|---|
| FormReply | AI auto-replies to form submissions. Full SaaS with OAuth and Stripe. | formreply.app |
| Changelog.dev | Beautiful changelogs for dev tools. GitHub integration + payments. | changelogdev.com |
| auto-co | This repo. The framework improving itself. | runautoco.com |
Monitoring
make monitor # live cycle output make status # current state summary make health # check loop health make history # cycle history with costs make export # export all data as JSON
Dashboard: app.runautoco.com (or run your own from projects/dashboard/).
Safety
Hard limits that can never be overridden:
- No repo/project/service deletion
- No database resets or force push to main
- No credential leaks to public repos
- No spending without human approval
Everything else — creating repos, deploying services, writing code — is fair game.
vs. Other Tools
| auto-co | Agent frameworks (LangGraph, CrewAI) | Paperclip | |
|---|---|---|---|
| Core | Bash loop + Claude Code | Python SDK with abstractions | TypeScript + Postgres orchestrator |
| Setup | npx create-auto-co |
Install SDK, define graphs | npx paperclipai onboard |
| State | Git + markdown files | In-memory or custom stores | PostgreSQL |
| Dependencies | Claude Code only | Framework + LLM provider | Node.js + PostgreSQL |
| Output | Deployed products | Task completions | Managed agent sessions |
| Complexity | ~50 lines of core logic | Medium-high | Medium |
The insight: Claude Code already handles tool use, code generation, and multi-step reasoning. You don't need a framework on top of an AI that can already code. You just need a loop and a shared notepad.
Project Structure
auto-co-meta/
├── auto-loop.sh # The loop
├── PROMPT.md # System prompt for each cycle
├── Makefile # All commands
├── .env.example # Configuration
├── memories/
│ └── consensus.md # Shared state between cycles
├── .claude/
│ ├── agents/ # 14 agent persona files
│ └── skills/ # Agent capabilities
├── templates/ # Starter templates (saas, docs-site, api-backend)
├── projects/ # Output: products built by auto-co
├── state/ # Structured logs (JSONL)
└── logs/ # Cycle logs
Cost
- Per cycle: ~$1.80 (Opus) or ~$0.50 (Sonnet)
- Infrastructure: ~$5-7/mo on Railway (optional)
- Typical project: 50-100 cycles to reach a deployed MVP = $90-180
License
MIT — see LICENSE
Built by an autonomous AI company. For autonomous AI companies.