WhatsApp에서 Claude에게 복사하여 붙여넣는 중이었습니다. 그래서 수정했습니다.
hackernews
|
|
📦 오픈소스
#ai
#ai 딜
#chatgpt
#claude
#gemini
#개발자
#메신저
#자가호스팅
#클로드
원문 출처: hackernews · Genesis Park에서 요약 및 분석
요약
개발자를 위해 로컬 환경에서 구동되는 오픈소스 기반 개인 AI 비서 'Deus'가 공개되었습니다. 이 도구는 왓츠앱, 텔레그램 등 다양한 메신저와 연동되어 음성, 이미지, 유튜브 영상까지 처리하며, 모든 대화를 기억하고 문맥을 파악하는 시맨틱 메모리 기능을 제공합니다. 특히 대화 주제를 자동으로 분류해 스스로 답변 품질을 개선하는 자가 발전 루프를 탑재했으며, 리눅스 컨테이너 기반의 샌드박스로 개인정보를 안전하게 보호하는 것이 특징입니다.
본문
A personal AI assistant that lives in your messaging apps, remembers everything, and gets smarter over time. Built for developers who want a private, self-hosted assistant with container isolation, semantic memory, and a self-improvement loop — all running locally on your machine. - Memory — Remembers everything across all your conversations. Ask it something you discussed weeks ago and it'll recall it precisely, using semantic search to find the most relevant context. - Messaging apps — WhatsApp, Telegram, Slack, Discord, and Gmail — each a standalone MCP package you install as needed. Switch between them freely — memory and context follow you everywhere. - Voice — Send a voice message and it transcribes and responds. Runs locally on Apple Silicon — nothing leaves your machine. - Vision — Send a photo or screenshot and it sees and responds to it. - Calendar — Reads and creates Google Calendar events. Ask what's on your schedule, or tell it to book something. - Web & video — Fetch YouTube transcripts, summarize videos, or browse the web — all from a chat message. - Scheduled tasks — Set it to do things automatically on a schedule (daily summaries, weekly recaps, reminders). - Self-improvement — Scores its own responses over time and learns from both failures and successes. Low-scoring responses generate corrective reflections; high-scoring ones extract positive patterns. Uses DSPy to optimize its own system prompt per domain once enough samples accumulate. - Domain detection — Automatically tags conversations by topic (marketing, engineering, study, writing, strategy) so the self-improvement loop can learn per-domain patterns and optimize accordingly. - Sandboxed & secure — Every conversation runs in an isolated Linux container. The AI can't access your host system beyond what you explicitly allow. - External projects — Run deus in any project directory to get a coding agent with your full Deus memory and preferences. Or register a project and work on it through your messaging apps — Deus mounts it into an isolated container, auto-detects the tech stack, and shadows sensitive files automatically. | Feature | Deus | OpenClaw | NemoClaw | Cortex | |---|---|---|---|---| | Persistent memory | Semantic + tiered | Basic hooks | Via OpenClaw | Neuroscience-grounded | | Self-improvement | DSPy per-domain eval loop | No | No | No | | Cross-session recall | Yes (sqlite-vec embeddings) | No | No | Yes | | Container isolation | Yes (Linux VM per group) | No | No | No | | Multi-channel | WhatsApp, Telegram, Slack, Discord, Gmail | No | No | No | | Token efficiency | ~1,800 tokens at session start | Higher | Higher | Not measured | | Local-first | Yes | Yes | Yes | Yes | | Setup time | ~5 min (claude /setup ) | ~15 min | ~20 min | ~10 min | - macOS (Apple Silicon recommended), Linux, or Windows - Claude Code installed and authenticated - Docker Desktop — the installer handles WSL 2 on Windows automatically - Node.js 20+, Python 3.11+ - A Gemini API key (free tier is enough) git clone https://github.com/sliamh11/Deus.git cd Deus claude Inside the Claude Code prompt: /setup # Install deps, configure runtime, build container, onboard /setup includes a Personality Kickstarter at the end: choose from curated behavioral bundles (universal defaults, developer workflow, student/learner mode), pick individual behaviors à la carte, and optionally seed the self-improvement loop with battle-tested reflections so it isn't starting cold. Switching from another AI? Give Deus a head start. Paste this prompt into your current AI (ChatGPT, Gemini, etc.) and send the output to Deus in your first conversation: I'm switching to a new AI assistant called Deus. Generate a structured summary about me that I can give it so it knows me from day one. Include: 1. **About me** — name, role/occupation, location, languages I speak 2. **What I use AI for** — the main topics and tasks I come to you with 3. **Communication style** — how I like responses (concise vs detailed, formal vs casual, code-heavy vs explanatory) 4. **Preferences** — things I've corrected you on or asked you to do differently 5. **Key context** — ongoing projects, goals, or background that shapes our conversations Be specific and factual. Skip anything generic. Format as plain text. A fresh clone has zero channels — you add only the ones you need: /add-whatsapp # Scan QR code to connect WhatsApp /add-telegram # Paste bot token to connect Telegram @Deus what's on my calendar tomorrow? @Deus summarize the YouTube video at @Deus remind me every Monday morning what I worked on last week - One Node.js process on the host. No microservices. - Each conversation group runs in its own container with an isolated filesystem. - Domain detection tags conversations by topic so the self-improvement loop learns per-domain patterns. | Command | What it does | |---|---| /compress | Save the current session to the vault and update the semantic index | /resume | Load core memory + warm tier (last 3 se
Genesis Park 편집팀이 AI를 활용하여 작성한 분석입니다. 원문은 출처 링크를 통해 확인할 수 있습니다.
공유