Lazynext — The Autonomous AI Company OS
A production, self-running organization of AI agents that builds, deploys, markets, sells, and supports a real software product with zero human intervention after setup. The fleet currently operates the Accessibility Checker — a live WCAG scanning product at checker.lazynext.com — handling the full lifecycle from free scans through reports, upgrade checkout, billing, email sequences, and support.
The entire backend runs on Cloudflare's serverless platform (Workers, D1, Workers KV, Vectorize, Pages, Containers, Cron) and Cloudflare Workers AI for LLM inference — no local services required beyond the agent runner.
Live platform
| Surface | URL | What it is |
|---|
| Marketing site | lazynext.com | Company site + status page |
| Product app | checker.lazynext.com | Accessibility Checker — scan UI, reports, checkout (installable PWA) |
| Product API | api.lazynext.com | Scan/report/badge/checkout API + /mcp + /a2a + widget.js |
| Founder dashboard | dashboard.lazynext.com | Next.js ops console (passphrase + optional TOTP; installable PWA) |
| Platform worker | ai-company-os.dry-hall-6a50.workers.dev | Internal state API — token-gated (returns 401 unauthenticated) |
| Status | status.lazynext.com | Redirects to lazynext.com/status |
The product — Accessibility Checker
The company's first product is a WCAG accessibility scanner (products/accessibility-checker/, mirrored at github.com/Lazynext-AI/accessibility-checker):
- Scan → report → upgrade loop: free tier 3 scans/day per site; full reports with per-rule findings; report pages carry the upgrade CTA into checkout
- 75-rule manifest (
GET /rules): WCAG 2.x criteria + best-practice checks, including interactive rendered-scan checks — real Tab/Shift+Tab keyboard-trap traces, Escape probes, click-opened dialog traps, target-size, focus-obscured, and non-text-contrast detection
- Billing: Dodo Payments — permanent test mode (founder decision 2026-09-29; no live KYC/webhook migration) — Pro subscription, 14-day trial (extendable via
config:trial_offer, currently 30 days ON), WELCOME20 discount (basis-points 2000 = 20%), webhook-verified entitlement, SignWell legal signing on activation (template unset, mechanism exists)
- Agent surfaces:
POST /mcp (JSON-RPC tools scan_url/scan_html/get_report/list_rules), POST /a2a (tasks/send/tasks/get), GET /.well-known/agent.json, GET /widget.js (Shadow-DOM embed) — all sharing the same quota-enforced scan pipeline
- SDKs:
sdk/js (@lazynext/accessibility-checker), sdk/go, sdk/python (with python -m CLI: scan/site/report/csv/rules/badge/card)
- Email: Brevo sequences + waitlist confirms (HMAC-verified unsubscribe) + engagement tracking feeding CRM campaigns
Table of Contents
- Live platform
- The product — Accessibility Checker
- Overview
- System Architecture
- Agent Types and Roles
- Workflow and Message Flow
- Reward and Performance System
- Technical Stack
- Installation
- Configuration
- Running the System
- Project Structure
- Performance and Scaling
- Production Deployment
Overview
Lazynext is a multi-agent system that runs a complete startup organization. Specialized AI agents — CEO, CTO, engineering, growth, and infrastructure roles — collaborate through a D1-backed event bus to operate the company's product end to end: shipping code, deploying to Cloudflare, drafting and sending campaigns, handling support, reconciling billing, and monitoring production. The fleet runs under launchd on a Mac runner with hard safety gates (protected files, test gate, fitness review, revert-on-failure, dead-corpus dedup) documented in AGENTS.md.
Key Capabilities
- Autonomous Product Development: Agents generate, validate, test, and commit code automatically
- Strategic Planning: CEO agent sets strategic direction based on company state and market analysis
- Task Orchestration: CTO agent decomposes strategic goals into executable technical tasks
- Code Generation: Engineering agents write production-quality code with validation and testing
- Automatic Deployment: DevOps agents deploy frontends to Cloudflare Pages and JS/TS backends to Cloudflare Workers, with verification
- Quality Assurance: Continuous testing and health monitoring with automatic remediation
- Performance Tracking: Comprehensive scoring and reward system for agent improvement
- Knowledge Management: RAG-powered knowledge base for context-aware decision making
Core Principles
- Separation of Concerns: Product code lives in a separate repository (
./products/<slug>/, e.g. products/accessibility-checker) isolated from the agent system
- Event-Driven Architecture: D1-backed message streams ensure reliable, exactly-once message delivery
- Tiered Model Selection: Cost-optimized LLM usage based on task importance
- Fail-Safe Operations: Automatic retries, rollbacks, and escalation protocols
- Production-Ready: Code validation, file backups, conflict detection, and git branch management
System Architecture
API Layer (Cloudflare Worker)
All state access goes through a single Cloudflare Worker (worker/) that fronts D1, KV, and Vectorize. Agents and the dashboard authenticate with a shared bearer token (API_TOKEN Worker secret).
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
