CategoryWeightCategory scoreContribution
Securityprompt, exec, net, exfil, eval
35%
35.0 pts
Supply chainhash, typosquat, maintainer, lockfile
20%
20.0 pts
Maintenancestaleness, pinning, CI
15%
15.0 pts
TransparencySKILL.md, perms, README
15%
15.0 pts
Communityinstalls, verify, response
15%
15.0 pts
Findings & checks · 0 flagged
Security
✓—All security checks passedNo findings in this category for the latest scan.pass
Supply chain
✓—All supply chain checks passedNo findings in this category for the latest scan.pass
Maintenance
✓—All maintenance checks passedNo findings in this category for the latest scan.pass
Transparency
✓—All transparency checks passedNo findings in this category for the latest scan.pass
Community
✓—All community checks passedNo findings in this category for the latest scan.pass
Every scanned point with the score it earned and what moved between them.
f02fcfc
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Stop Paying Claude to Write Boilerplate
<p align="center"> <img src="./landing/promo.gif" alt="4AI Collaborative Development Demo" width="700"> </p>
<p align="center"> Claude designs. Codex implements for <strong>$0</strong>. Gemini researches for <strong>free</strong>.<br> A CLI that routes each task to the right AI — cutting costs by 75%. </p>
<p align="center"> <a href="./README_ja.md">日本語</a> · If this project helps you, please <a href="https://github.com/AI-Driven-School/aiki/stargazers">give it a star</a> </p>
Get Started in 30 Seconds
# Option A: npx (recommended)
npx aiki init my-app
# Option B: curl
curl -fsSL https://raw.githubusercontent.com/AI-Driven-School/aiki/main/install-fullstack.sh | bash -s -- my-app
# Then start developing
cd my-app && claude
# Build a feature
> /project user authenticationRequirements:
| AI | Role | Cost |
|---|---|---|
| Claude Code | Design & Review | Pay-per-use |
| Codex (ChatGPT Pro) | Implementation & Testing | Included in subscription |
| Gemini CLI | Large-scale analysis | Free |
| Grok (xAI) | Real-time trends & X search | xAI API |
Why 4 AIs?
┌─────────────────────────────────────────────────────────────┐
│ │
│ Single AI 4-AI Collaboration │
│ ───────── ────────────────── │
│ Claude only Claude → Design & decisions only │
│ = Everything billed ↓ │
│ = High cost Codex → Implementation & tests ($0) │
│ ↓ │
│ Claude → Review │
│ │
│ Result: $1.00 Result: $0.25 (75% reduction) │
│ │
└─────────────────────────────────────────────────────────────┘Benchmark Results
| Task Type | Winner | Reason |
|---|---|---|
| New file creation | Codex | 10-20% faster, $0 |
| Existing code modification | Claude | 40% faster, 2x quality |
| Large-scale analysis | Gemini | 1M tokens, free |
| Real-time trends | Grok | Live X data, xAI API |
Detailed benchmark results
Workflow
> /project user authentication
[1/6] Requirements (Claude) → docs/requirements/auth.md ✓
[2/6] API Design (Claude) → docs/api/auth.yaml ✓
[3/6] Implementation(Codex) → src/**/*.tsx ★ full-auto
[4/6] Testing (Codex) → tests/*.spec.ts ★ $0
[5/6] Review (Claude) → Improvement suggestions ✓
[6/6] Deploy → https://my-app.vercel.app ✓Output
my-app/
├── .rulesync/rules/ # Single source of truth for all AI rules
├── docs/
│ ├── requirements/ # Requirements (Claude)
│ ├── specs/ # UI specs (Claude)
│ └── api/ # API design (Claude)
├── src/ # Implementation (Codex)
└── tests/ # Tests (Codex)Even solo developers get comprehensive design documentation.
Case Study: StressAIAgent
A real SaaS built with this template:
AI-Powered Mental Health SaaS
/project organization analysis AI| Phase | AI | Output |
|---|---|---|
| Requirements | Claude | docs/requirements/org-analysis-ai.md |
| API Design | Claude | docs/api/org-analysis.yaml |
| UI Specs | Claude | docs/specs/org-analysis-ai.md |
| Implementation | Codex | Backend + Frontend |
| Testing | Codex | 10 E2E test cases |
| Review | Claude | docs/reviews/org-analysis-ai.md |
Generated Features
- Organization-wide stress analysis dashboard
- Department score heatmap
- GPT-4 powered AI insights generation
- PDF report export
- Admin permission checks
Source code
Real-world Case: Multi-Agent Debate caught a critical legal misinterpretation
A 2026-05 case study where Solver/Proposer/Checker debate found a 致命的な事実誤認 in 1 job (~2 minutes):
- Solver (Claude) claimed a SaaS subscription violated Japan's 特商法 (Specified Commercial Transactions Act)
- Proposer (Gemini) rebutted: 特商法「特定継続的役務」is a 政令で7業種限定列挙, SaaS is not included
- Checker (Claude+WebSearch) confirmed via 消費者庁 official guide
ROI: ~2min debate avoided shipping a wrong legal interpretation in client materials.
Read the full case study
Benchmark (2026-05-08): 4-agent debate at n=30 — Recall 100%, F1 87.8%
Tested with Solver / Proposer / Critic / Checker pattern using 4 different model families:
| Approach | Recall (false catch) |
|---|---|
| Single-AI review | 44–54% |
| Greptile (commercial best) | 82% |
| MARCH/CORE research target | 85–90% |
| This stack — Stage 1 (n=6) | 100% (6/6) |
| This stack — Stage 2 (n=30) | 100% (18/18 false caught) / F1 87.8% |
Per-reviewer catch rate (Stage 2 false-claim subset, n=18):
- Codex Critic alone (
-m gpt-5.5, 88.7% SWE-Bench): 89% (16/18) — strongest single reviewer - Gemini Proposer alone (3.x): 78% (14/18)
- Combined via debate: 100% (18/18)
Caveats:
- n=30 still has ±5–10pp confidence interval. Selection bias and self-evaluation present.
- Greptile/PR-Agent/Qodo numbers (82%, 44%, 60.1%) are for PR-diff bug catch — not directly comparable. Our /debate evaluates natural-language claims (business judgment, legal interpretation, technical assertions). Different ground. We tried direct comparison with PR-Agent and confirmed it requires PR URLs as input — see vs PR-Agent comparison.
- Independent grader confirmed Recall 100% holds with κ=0.534 moderate agreement on verdict classifications — see independent grading report.
Stage 3 plan (n=100, independent grader, 20 code-PR-format claims for direct PR-Agent comparison).
Codex -m gpt-5.5 brought primary sources Gemini missed (PR TIMES, Meta Threads API spec, Vercel pricing). Vendor-independent rebuttals across Anthropic + Google + OpenAI + WebSearch are real, not theoretical.Commands
| Command | AI | Description |
|---|---|---|
/project <feature> | All | Complete flow: design → implementation → deploy |
/requirements <feature> | Claude | Generate requirements |
/spec <screen> | Claude | Generate UI specifications |
/api <endpoint> | Claude | Generate API design |
/implement | Codex | Implement from design docs |
/test | Codex | Generate tests |
/review | Claude | Code review |
/analyze | Gemini | Large-scale code analysis |
/research <query> | Gemini | Technical research |
/x-trend-research | Grok | X trend search |
/x-context-research | Grok | Context pack for article writing |
Claude Code Skills (13 Skills Included)
This project ships with 13 ready-to-use Agent Skills for Claude Code.
Install Skills
# Using Vercel skills CLI
npx skills add AI-Driven-School/aiki
# Or clone and use directly
git clone https://github.com/AI-Driven-School/aiki.git
cd aiki && claudeAvailable Skills
| Skill | AI | Description |
|---|---|---|
/project | All | Full design-to-deploy pipeline (6 phases) |
/requirements | Claude | Generate requirements with user stories & acceptance criteria |
/spec | Claude | Generate UI screen design documents |
/api | Claude | Generate OpenAPI 3.0 specifications |
/implement | Codex | Auto-generate code from design docs |
/test | Codex | Generate E2E & unit tests from requirements |
/review | Claude | Code review against design documents |
/analyze | Gemini | Full codebase analysis (1M token context) |
/research | Gemini | Technical research with comparison reports |
/frontend-design | Claude | Production-grade UI without generic AI aesthetics |
/mockup | Claude | Generate UI mockup PNGs via Playwright |
/mockup-swift | Claude | Native iOS/macOS mockups via SwiftUI Preview |
/checkpointing | Claude | Save & restore session state across sessions |
All skills follow the Agent Skills open standard and work with Claude Code, Codex CLI, OpenCode, Cursor, and other compatible agents.
Unified Rule Management (rulesync)
NEW: Manage all AI rules in one place. rulesync generates config files for each AI from a single source.
# Edit rules once
vim .rulesync/rules/project-overview.md
# Generate for all AIs
rulesync generate
# Output:
# → CLAUDE.md (Claude Code)
# → AGENTS.md (Codex CLI)
# → GEMINI.md (Gemini CLI)
# → .claude/rules/*.md (Claude delegation rules)
# → .codex/memories/*.md (Codex shared context)
# → .gemini/memories/*.md (Gemini shared context)| Source | Generated For | Description |
|---|---|---|
project-overview.md | All 3 AIs | Shared project context |
claude-workflow.md | Claude only | Enforcement rules, commands |
codex-agent.md | Codex only | Implementation guidelines |
gemini-agent.md | Gemini only | Research guidelines |
*-delegation.md | Claude only | When to delegate to each AI |
Grok is managed manually via .grok/GROK.md (no rulesync target yet)Auto-Orchestration (MCP Server)
Talk naturally to Claude, and tasks are automatically delegated to the right AI.
You: "Implement user authentication"
→ Automatically delegated to Codex (implementation specialist)
You: "Compare React vs Vue"
→ Automatically delegated to Gemini (research specialist)
You: "Review my code"
→ Handled by Claude (design & review specialist)Setup
# Install and configure MCP server
./scripts/setup-mcp.sh
# Restart Claude CodeNo API keys needed - Uses CLI tools (Codex CLI from ChatGPT Pro, Gemini CLI free).
Available MCP Tools
| Tool | Description |
|---|---|
delegate_to_codex | Send implementation tasks to Codex |
delegate_to_gemini | Send research/analysis to Gemini |
auto_delegate | Automatically route to the right AI |
get_orchestration_status | Check AI service status |
See MCP Server README for details
Adoption Paths
Start small and add AIs as needed:
| Mode | Command | AIs | Best for |
|---|---|---|---|
| Claude only | --claude-only | Claude | Getting started, small projects |
| Claude + Codex | --claude-codex | Claude, Codex | Implementation-heavy projects |
| Claude + Gemini | --claude-gemini | Claude, Gemini | Research-heavy projects |
| Full (default) | --full | All 4 | Maximum productivity |
# Start with Claude only
bash install-fullstack.sh my-app --claude-only
# Later, upgrade to full 4-AI
bash install-fullstack.sh my-app --fullCompatibility
AI CLI Support
| Tool | Min Version | Max Tested | Required |
|---|---|---|---|
| Claude Code | 1.0.0 | 2.1.49 | Yes |
| Codex CLI | 0.1.0 | 0.93.0 | No |
| Gemini CLI | 0.1.0 | 0.26.0 | No |
| Ollama | 0.1.0 | — | No |
Version compatibility is automatically checked during installation and delegation. See .ai-versions.json for details.
IDE & Agent Support
| Tool | Support | Notes |
|---|---|---|
| Claude Code | Native | Full skills, hooks, and delegation |
| Codex CLI | Native | Via AGENTS.md and delegation scripts |
| OpenCode | Compatible | Reads AGENTS.md and .claude/skills/ natively |
| Cursor | Compatible | Via .cursor/rules/ templates |
| VS Code | Tasks | Via templates/vscode-tasks.json |
See OpenCode Compatibility for details.
Local Model Support
Run the entire pipeline offline with Ollama:
# Code generation with local model
bash scripts/delegate.sh ollama generate "Create a REST API"
# Code review with local model
bash scripts/delegate.sh ollama review src/ --model deepseek-coderSystem Requirements
- macOS / Linux / WSL2
- Node.js 18+
- ChatGPT Pro (for Codex CLI, $200/month) - optional with
--claude-onlymode - Gemini CLI (free, optional for research tasks)
Documentation
| Document | Description |
|---|---|
| Getting Started | Installation & setup |
| Hands-on Tutorial | Build a TODO app step by step |
| Existing Project Setup | Add to existing projects |
| Cost Comparison | Detailed cost analysis: 4-AI vs single AI |
| OpenCode Compatibility | Using with OpenCode CLI |
| Command Reference | Complete command reference |
| Benchmarks | Detailed benchmark data |
Support
Phase 1 is completely free. The easiest way to support this project:
- Give it a star on GitHub - helps others discover this project
- Share on X/Twitter - spread the word
- Sponsor development - support ongoing work
<p align="center"> MIT License<br> <a href="https://github.com/AI-Driven-School/aiki">GitHub</a> · <a href="https://github.com/AI-Driven-School/aiki/issues">Issues</a> · <a href="https://github.com/sponsors/AI-Driven-School">Sponsor</a> </p>