Self-improving AI operating system powered by the Claude Agent SDK — no API credits, runs on your Claude Max account. 47 specialist agents with live tool access (filesystem, Bash, AWS CLI, web), persistent memory, cross-session goal tracking, and one-command AWS SaaS generation.
- Agent SDK engine — All agent calls route through
@anthropic-ai/claude-agent-sdk. Uses Claude Code Max account — zero API credits consumed. - Real tool access — Every agent can Read/Write files, run Bash (including
awsCLI), search the web. Agents don't just generate text — they act. - AWS supercharged —
balancedandpremiumprofiles include Bash + AWS CLI. The deployer agent provisions real infrastructure. - Claude API fallback — If Agent SDK is unavailable, falls back to direct
claude-opus-4-6with adaptive thinking. - No required API keys — Just needs Claude Code authenticated with Max account.
- 47 specialist agents across 10 departments + intelligence layer
- Persistent memory (episodic, semantic, procedural) with automatic context injection
- Cross-session goal tracking with deadline management and blocker escalation
- Multi-modal perception (vision, PDF, audio) with automatic input detection
- Self-improving reflection engine that scores outputs and extracts reusable strategies
- AWS SaaS generator — one command produces complete CDK infrastructure + Node.js backend
- Blockchain module — BTC, ETH, XRP, SOL
- REST API — 50+ endpoints, SSE streaming, web UI at
localhost:3000
git clone <your-repo-url>
cd Genie
npm installcp .env.example .env
# Add your ANTHROPIC_API_KEY — that's all you need# CLI
node src/index.js "build me a SaaS app for tracking freelance invoices"
# Web UI + API
npm run server
# open http://localhost:3000| Profile | Engine | Tools | Best for |
|---|---|---|---|
fast |
Agent SDK | Read, Glob, Grep | Formatting, lookups, simple ops |
balanced |
Agent SDK | + Write, Edit, Bash, WebSearch | Most agent tasks, AWS ops |
premium |
Agent SDK | All tools + WebFetch | Architecture, planning, full builds |
Fallback (if Agent SDK unavailable): direct Claude API (Opus 4.6 adaptive thinking).
node src/index.js "task" # standard run
node src/index.js --interactive "task" # step-by-step approval
node src/index.js --research "question" # research only, no code output
node src/index.js --power 3 "task" # power level 1-5
node src/index.js --file path/to/task.json # load task from fileCLI / HTTP API
→ workflow.js routes task, drives iteration loop
→ orchestrator.js multi-iteration agent execution
→ agents/*.js 47 specialist agents
→ claudeClient.js Claude API engine (Opus / Sonnet / Haiku)
→ patchExecutor.js writes output to my-workspace/output/
→ memoryManager.js injects persistent brain into every call
my-workspace/
output/ generated projects
reports/ research, PDFs, markdown
requests/ task JSON files
memory/
semantic/ learned facts & preferences (knowledge.json)
episodes/ run history
procedural/ proven strategies
# No API key required — uses Claude Code Max account auth
OPENAI_API_KEY= # optional — DALL-E image generation only
ANTHROPIC_API_KEY= # optional — fallback if Agent SDK unavailable
AWS_ACCESS_KEY_ID= # optional — AWS deployment agent
AWS_SECRET_ACCESS_KEY=
AWS_REGION=
MAX_ITERATIONS=5
PORT=3000
- Copy
src/agents/core/AGENT_TEMPLATE.jsinto the right department folder - Implement
orchestrate()(orrun()for simple agents) - Import + instantiate in
src/index.jsandsrc/server.js - Register in
src/agentRegistry.js
Agents call Claude via:
import { consensusCall } from "../../llm/multiLlmSystem.js";
const result = await consensusCall({
profile: "premium", // fast | balanced | premium
system: "...",
user: "...",
schema: { name: "output", schema: { ... } } // optional — returns parsed JSON
});docker-compose -f docker-compose.genie.yml up -d --build
docker-compose -f docker-compose.genie.yml down
docker logs genie-server