ZeusHammer – “국부적으로 사고하는” AI 에이전트 개발
hackernews
|
|
📦 오픈소스
#anthropic
#claude
#gpt-4
#llama
#openai
원문 출처: hackernews · Genesis Park에서 요약 및 분석
요약
ZeusHammer는 세 가지 오픈 소스를 기반으로 한 로컬 사고형 AI 에이전트로, 의도 인식과 스킬 매칭의 80%를 로컬에서 처리해 비용을 절감하고 응답 속도를 높였습니다. 웨이크워드부터 TTS까지 지원하는 음성 우선 설계와 세 가지 계층의 메모리 시스템을 갖추고 있으며, 유휴 시간에 자가 진화하는 명상 모드와 스킬을 자동 생성하는 워크플로우 엔진을 제공합니다.
본문
The Super AI Agent Built from Three Open Source Giants Think Locally. Speak Freely. Remember Everything. | Feature | OpenClaw | Hermes | ZeusHammer | ZeusHammer Advantage | |---|---|---|---|---| | Local Brain | ❌ | ❌ | ✅ | 80% tasks without LLM, intent recognition + skill matching | | Voice-First | Partial (Voice Wake) | Partial | ✅ | Whisper+Edge TTS full-stack, wake word, auto language detection | | Three-Tier Memory | ❌ | ❌ | ✅ | Short-term (LRU)+Long-term (SQLite)+Working memory | | Meditation Mode | ❌ | ❌ | ✅ | Auto analysis/pattern extraction/skill upgrade on idle | | Deep Reflection | ❌ | ❌ | ✅ | Causal chain analysis/insight generation | | Workflow Engine | Standing Orders | Cron/SubAgent | ✅ | Complete orchestration+skill matching+skill learning | | Circuit Breaker | ❌ | ✅ Circuit breaker | ✅ | Complete security system | | OSV Scanning | ❌ | ❌ | ✅ | Command/file malware detection | | China LLM Native | Needs config | Partial | ✅ | Built-in DeepSeek/Qwen/GLM/Moonshot/ERNIE/Doubao/MiniMax | | Tool Detector | ❌ | ❌ | ✅ | Auto-detect 80+ system CLI tools | | Cost Tracking | ❌ | Partial | ✅ | Built-in LLM cost tracking | | Smart Model Router | ✅ | ✅ | ✅ | Intelligent routing+fallback chain+cost optimization | | WebRTC Real-time Audio | ❌ | ❌ | ✅ | Real-time voice streaming | | Collaboration Rooms | ❌ | ❌ | ✅ | Multi-user collaboration | | Apple Ecosystem | ✅ | ✅ | ✅ | Apple Notes/Reminders/iMessage/Screenshot | | Vector Memory Store | LanceDB/QMD | External plugins (8) | ✅ | Built-in Chroma, native integration | | Skill Auto-Generation | ❌ | ❌ | ✅ | Local Brain auto-generates new skills | | RL Training | ❌ | ✅ Tinker-Atropos | ✅ | SWE-Bench/RL/trajectory compression | | Install Convenience | npm/global | pip | Homebrew/pipx/Docker | Best for macOS users | | Message Channels | 30+ | 20+ | 13+ | OpenClaw has most | | Built-in Tools | 20+ | 40+ | 50+ | ZeusHammer has most | | AI Models | 35+ | 15+ | 20+ | OpenClaw has most | | Skills | Plugin market | 70+ | 50+ | Hermes has most | | MCP Support | ✅ | ✅ | ✅ | All three have | | ACP/IDE Integration | ✅ | ✅ | ✅ | All three have | | Browser Automation | ✅ | ✅ | ✅ | All three have | | Community Scale | 400+ contributors, ClawHub | Nous Research support | New project | OpenClaw most mature | ZeusHammer has the following unique differentiators compared to OpenClaw and Hermes: - Local Brain Architecture — The biggest differentiator. 80% of tasks complete via intent recognition + skill matching locally, without calling expensive LLM APIs, significantly reducing costs and improving response speed. - Voice-First Design — Complete voice-first experience from wake word, STT, TTS to auto language detection, supporting background Daemon mode for continuous listening. - Three-Tier Memory System — Short-term (LRU) + Long-term (SQLite FTS5) + Working memory, plus vector storage (Chroma), providing the most complete memory capabilities. - Meditation Mode + Deep Reflection — Industry-rare self-evolution capability. AI automatically analyzes memory patterns, extracts rules, and upgrades skills during idle time. - Workflow Engine — Includes skill matching and skill learning. Not just scheduling existing skills, but also auto-generating new skills. - Security System — Four layers: OSV malware scanning, circuit breaker, credential guard, and config protection. - Tool Detector — Auto-discovers ~80 common CLI tools in the system, no manual configuration needed. - China LLM Native Support — Built-in unified clients for Chinese mainstream models like DeepSeek/Qwen/GLM/Moonshot/ERNIE/Doubao/MiniMax, accessible via chinawhapi.com. - Homebrew Installation — macOS users can install with one command: brew install zeushammer , providing the best user experience. - 50+ Built-in Tools — More out-of-box tools than OpenClaw and Hermes combined. Weaknesses: Ecosystem scale (contributors, skill marketplace) not as mature as OpenClaw; message channels (13+) less than OpenClaw (30+) and Hermes (20+); as a new project, community and documentation maturity need verification. ZeusHammer is a true fusion of three top-tier open source AI agent projects, combining their best features into one unified super agent: | Project | Core Contribution | License | |---|---|---| | ClaudeCode | Tool execution engine, concurrent partitioning, OTel telemetry | Anthropic | | Hermes | Memory system, security, tools, MCP protocol | MIT | | OpenClaw | Multi-channel gateway, Canvas, skills management | MIT | partitionToolCalls() - Concurrent partitioning algorithmisConcurrencySafe() - Concurrency safety checksToolResult - Unified tool result format- OTel structured logging and tracing - Memory System: Short-term + Long-term + Vector memory with FTS5 search - Security: OSV malware scanner, credential guard, circuit breaker - Tools: Tool registry, MCP protocol stack, skill auto-improvement - Skills: Agent-curated skill creation and management - Multi-Channel: Telegram, Discord, Slack, WhatsApp gateway - Canvas: Live Canvas rendering system - Skills: SkillHub with marketplace integration - Config: Protected paths and tamper-proof configuration brew install zeushammer Or with cURL: curl -sSL https://raw.githubusercontent.com/pengrambo3-tech/ZeusHammer/master/install.sh | bash Or manual install: git clone https://github.com/pengrambo3-tech/ZeusHammer.git cd ZeusHammer pip install -r requirements.txt ┌─────────────────────────────────────────────────────────────────┐ │ ZeusHammer │ ├─────────────────────────────────────────────────────────────────┤ │ ┌───────────────────────────────────────────────────────────┐ │ │ │ Fusion Core │ │ │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────┐ │ │ │ │ │ClaudeCode │ │ Hermes │ │ OpenClaw │ │ │ │ │ │────────────│ │────────────│ │────────────────────│ │ │ │ │ │Tool Engine │ │ Memory │ │ Channel Gateway │ │ │ │ │ │Partition │ │ Security │ │ Canvas System │ │ │ │ │ │Telemetry │ │ MCP │ │ Skills Hub │ │ │ │ │ └─────────────┘ └─────────────┘ └─────────────────────┘ │ │ │ └───────────────────────────────────────────────────────────┘ │ │ ┌───────────────────────────────────────────────────────────┐ │ │ │ Local Brain │ │ │ │ Intent Recognition → Skill Match → LLM Work → Learn │ │ │ └───────────────────────────────────────────────────────────┘ │ │ ┌───────────────────────────────────────────────────────────┐ │ │ │ Voice System │ │ │ │ Wake Word "Zues/宙斯" → STT → Language → TTS │ │ │ └───────────────────────────────────────────────────────────┘ │ │ ┌───────────────────────────────────────────────────────────┐ │ │ │ Three-Tier Memory │ │ │ │ Short-term (LRU) → Long-term (SQLite) → Working │ │ │ └───────────────────────────────────────────────────────────┘ │ ├─────────────────────────────────────────────────────────────────┤ │ LLM Providers: ChinaWhapi • Anthropic • OpenAI • Local │ └─────────────────────────────────────────────────────────────────┘ The fusion brings together the most powerful features from each project: | Feature | From | Description | |---|---|---| | Tool Partition | ClaudeCode | Concurrent tool execution with dependency analysis | | Circuit Breaker | Hermes | Fail-fast pattern for stable operation | | Multi-Channel | OpenClaw | Connect via Telegram, Discord, Slack, etc. | | OSV Scanner | Hermes | Detect malware patterns in code | | Credential Guard | Hermes | Prevent credential leakage | | Memory System | Hermes | Persistent cross-session memory | | Skill Learning | Hermes | Auto-create skills from experience | | Canvas | OpenClaw | Live interactive canvas | | OTel Telemetry | ClaudeCode | Structured logging and tracing | # ZeusHammer's Local Brain workflow: 1. User gives instruction 2. Local Brain receives instruction 3. Pattern matching against skill library 4. Match found → Execute skill directly (NO LLM needed!) 5. No match → Call LLM for solution 6. Work complete → Learn new skill 7. Next time → Use learned skill (instant, no LLM) | Component | Technology | Benefit | |---|---|---| | Speech-to-Text | Whisper (local) | Offline capable, no data sent | | Text-to-Speech | Edge TTS | Natural, free voices | | Wake Word | Custom detector | "Zues" or "宙斯" activation | | Language Detection | Auto-detection | Auto-switch Chinese/English | | Smart Responses | Context-aware | Reply based on model/memory status | | Provider | Models | Features | |---|---|---| | DeepSeek | V3, Coder | Best value, coding focused | | Qwen (Alibaba) | Turbo, Plus, Max | Long context support | | GLM (Zhipu) | GLM-4, GLM-4V | Vision support | | Moonshot | 8K, 32K, 128K | Ultra long context | | ERNIE (Baidu) | Bot 4.0, Bot Long | Enterprise grade | | Doubao (ByteDance) | Pro, Lite | Fast, cost effective | | MiniMax | ABAB6 Chat/GSPT | Conversational AI | | Provider | Models | |---|---| | Anthropic | Claude 3.5 Sonnet, Opus, Haiku | | OpenAI | GPT-4o, GPT-4 Turbo, GPT-3.5 | | Local | Ollama, LM Studio, vLLM | # Option 1: ChinaWhapi (recommended for Chinese users) echo "OPENAI_API_KEY=your_key" >> ~/.zeushammer/.env echo "API_BASE=https://api.chinawhapi.com/v1" >> ~/.zeushammer/.env echo "MODEL=deepseek-chat" >> ~/.zeushammer/.env # Option 2: Anthropic Claude echo "ANTHROPIC_API_KEY=sk-ant-xxx" >> ~/.zeushammer/.env echo "MODEL=claude-3-5-sonnet-20241022" >> ~/.zeushammer/.env # Option 3: OpenAI echo "OPENAI_API_KEY=sk-xxx" >> ~/.zeushammer/.env echo "MODEL=gpt-4o" >> ~/.zeushammer/.env python3 -m src.main --mode cli # CLI mode python3 -m src.main --mode web # Web UI python3 -m src.main --mode voice # Voice (recommended!) | Level | Description | |---|---| safe | All operations require confirmation | semi_open | Safe operations auto-execute, dangerous operations warn | full_open | Unrestricted (beast mode) | Built-in protections (from Hermes): - OSV malware pattern scanning - Credential leakage detection - Circuit breaker for abnormal operations - Protected paths from OpenClaw - Config tamper protection # Run tests pytest tests/ # Code format black src/ ruff check src/ ZeusHammer/ ├── src/ │ ├── fusion/ # Three project fusion │ │ ├── claude_code/ # ClaudeCode core │ │ │ └── tools_engine.py │ │ ├── hermes/ # Hermes core │ │ │ └── security.py │ │ └── openclaw/ # OpenClaw core │ │ └── channels.py │ ├── brain/ # Local brain │ ├── voice/ # Voice system │ ├── memory/ # Memory system │ └── tools/ # Tools ├── FUSION.md # Fusion documentation ├── README.md └── README_zh.md Issues and Pull Requests welcome! MIT License Built with ❤️ by fusing ClaudeCode + Hermes + OpenClaw [Think Locally. Speak Freely. Remember Everything.]
Genesis Park 편집팀이 AI를 활용하여 작성한 분석입니다. 원문은 출처 링크를 통해 확인할 수 있습니다.
공유