Ellory 2.0

Autonomous AI engineering pipeline that turns a product idea into a runnable Next.js + shadcn/ui codebase. Multi-agent LangGraph graph with PostgreSQL persistence.

Stack

Generated apps: Next.js 16 (App Router, React 19, TypeScript, Turbopack), Tailwind CSS 4, shadcn/ui v4
Pipeline: FastAPI + LangGraph, PostgreSQL (AsyncPostgresSaver), Context7 MCP for live docs

Project Structure

ellory-next/
├── apps/
│   ├── web/               # Next.js 16 UI (idea submission, SSE streaming, project timeline)
│   └── backend/           # FastAPI + LangGraph pipeline
│       ├── app/
│       │   ├── core/      # settings, logger, llm client, checkpointer
│       │   ├── graph/     # builder.py, state.py, nodes/, prompts/
│       │   ├── db/        # SQLAlchemy models + service layer
│       │   ├── routers/   # API endpoints
│       │   ├── services/  # GitHub + git integrations
│       │   └── tools/     # fs, shell, workspace, mcp tools
│       └── migrations/    # Alembic DB migrations
├── docker-compose.yml
└── turbo.json

Pipeline Nodes

1. CEO — refines raw idea, scopes project
2. Decision Engine (×3 max) — strategic/technical decisions (auth, persistence, etc.)
3. Researcher — parallel Context7 queries for live stack docs
4. Architect ↔ Critic (≥9/10 to pass, ×3 max) — architecture design + critique
5. Planner — breaks architecture into ordered build phases
6. GitHub Setup — creates repo (skipped if no GITHUB_TOKEN)
7. Stack Researcher — pulls final package versions + config patterns
8. Design Director — establishes visual design language
9. Stack Bootstrapper — runs create-next-app, shadcn init, vitest setup (no LLM)
10. Test Writer — writes failing tests before implementation
11. Agentic Builder — ReAct agent implements each phase using shell + Context7
12. Consistency Fixer — tsc --noEmit + pnpm lint loop (≤3 passes)
13. Code Reviewer (≥8/10 to pass, ×3 max) — security, startup blockers, best practices
14. Docs Writer — generates README.md + AGENTS.md in workspace
15. Product Manager — decides next feature; exits when all shipped
16. Feature Test Writer / Feature Builder — TDD loop per feature → back to Consistency Fixer

Getting Started

Prerequisites

• Node.js ≥ 18 + pnpm
• Python ≥ 3.12 + uv
• Docker
• LLM API key (OpenAI, Anthropic, or any LiteLLM-compatible provider)
• Context7 API key

1. Start Postgres

docker compose up -d

2. Configure the backend

cp apps/backend/.env.example apps/backend/.env

Minimum required vars:
LLM_MODEL and corresponding API key variable can be found at https://models.litellm.ai/

LLM_MODEL=
<MODEL_API_KEY>
CONTEXT7_API_KEY=ctx7sk-...

Optional (enables GitHub repo creation):

GITHUB_TOKEN=ghp_...
GITHUB_OWNER=your-github-username

3. Start the backend

cd apps/backend
uv sync
uv run uvicorn app.main:app --port 8000 --env-file .env

4. Start the frontend

cd apps/web
pnpm install
pnpm dev

Open http://localhost:3000.

API

Method	Endpoint	Description
POST	/generate	Synchronous run
POST	/generate/stream	SSE streaming (node completions, tool calls, logs)
POST	/projects/{id}/resume/stream	Resume from last checkpoint
GET	/projects	Project list
GET	/projects/{id}	Project detail + artifacts
GET	/projects/{id}/features	Feature list
GET	/projects/{id}/checkpoint	Raw LangGraph checkpoint state

Persistence

• PostgreSQL — orchestration state: decisions, architecture, critic scores, revision counters, node progress
• Disk ($WORKSPACE_ROOT/<project_id>/) — generated file contents (source of truth)