daozhu

道樞 Dàoshū — An Eastern-Philosophy AI Orchestrator

Languages: English · 繁體中文

道樞 (Dàoshū, "Pivot of the Tao") is a self-adaptive AI agent framework grounded in Eastern philosophy. A seven-layer perception network runs in the background to answer three questions about every input — who are you, how are you, what do you need — then decides which philosophical school should respond.

Feeling low? It answers as Taoism (道家). Stuck with a person? Confucianism (儒家). Negotiating? Strategic School (縱橫家). Torn between options? It convenes a cabinet meeting. Every reply is prefixed with 【感知：XXX】 (perception tag).

Features

• Seven-layer perception network — emotion, task, interpersonal, complexity, concise-mode, tone, energy — fully automatic in the background
• Cabinet of four schools — Confucian (dual internal/external), Taoist (three-tier), Legalist (dual-domain / three-table), Strategist (opening-three-questions)
• Three advisory patches — Military (embedded in Strategist), Mohist logic (in Legalist), Buddhist insight (in Taoist)
• Seven skill modules — Work Guide, Virtual Masters, Music Companion, Metaphysical Mirror (hexagram/tarot/astrology/feng-shui/archetype/dream/oracle), Solar Terms, Cabinet Meeting, Skill Routing (hook in your own skills)
• Memory layer MCP — 27 tools: expense, health, reminders, shopping, mood journal, study notes (incl. mark-reviewed), decision log, weekly report, energy insight, strategy vault (道藏), solar-term calendar, CSV export, config
• Local-first — memory stored in local_memory/ as JSON, atomic writes prevent corruption, nothing leaves your machine

Architecture

User input
    ▼
┌─────────────────────────────────────┐
│  Perception Network (7 layers)      │
│  emotion → task → interpersonal     │
│  → complexity → concise → tone      │
│  → energy                           │
└─────────────────────────────────────┘
    ▼
┌─────────────────────────────────────┐
│  Dispatch layer                     │
│  Cabinet (4 schools + 3 patches)    │
│  + 7 skill modules + routing        │
└─────────────────────────────────────┘
    ▼
┌─────────────────────────────────────┐
│  Memory layer (daozhu-mcp · 27 tools)│
└─────────────────────────────────────┘

Installation (Claude Code)

1. Copy the skill and MCP server into your project:

   cp -r skills/daozhu   <your-project>/.claude/skills/daozhu
   cp -r mcp             <your-project>/.claude/daozhu-mcp
   

2. Install Python dependencies (for the memory layer):

   pip install -r .claude/daozhu-mcp/requirements.txt
   

3. Configure MCP: copy .mcp.json.example to .mcp.json at your project root and edit paths.

4. Trust the project MCP servers — add to .claude/settings.local.json:

   { "enableAllProjectMcpServers": true }
   

   ⚠️ Note for VSCode extension environments: Claude Code reads the repo-root .mcp.json, not .claude/mcp.json.

5. Optional: reference it in your CLAUDE.md:

   All conversations are dispatched by the `daozhu` skill — see `.claude/skills/daozhu/SKILL.md`.
   

6. Reload Window, run /mcp, confirm daozhu shows Connected.

Quick Start

You say	道樞 does
"I'm so frustrated / exhausted"	Taoism + Buddhist patch, 【感知：情緒低落】
"How do I deal with my professor/teammate"	Confucian external + Strategist
"Should I take this project?" (complex)	Cabinet meeting (5 stages)
"Help me plan my study schedule"	Work Guide (4 phases)
"Lunch cost 150"	Expense log + auto-categorize
"I'm feeling down today"	Mood journal + classification + streak detection
"How was my week?"	Weekly report + energy insight
"System status"	Memory-store health check

Command Reference

Command	Function
/儒家 / /道家 / /法家 / /縱橫家	Switch school manually
/會議 [question]	Force a cabinet meeting
/工 [task] / /工 急 / /工 深 / /工 程式	Work Guide
/費曼 /蘇格拉底 /張愛玲 …	Summon a virtual master
/[musician] / /播 [mood]	Music companion
/卦 /塔羅 /星盤 /風水 /星 /夢 /緣	Metaphysical mirror (7 modes)
/節氣	Solar-term wellness reminder
/兵 /辯 /觀	Force Military/Mohist/Buddhist patch
/外治 /內修 /域	Legalist domain switch
/藏	Save a strategy to the vault
/快 / /慢	Toggle concise mode

Routing to Your Own Skills

When the perception network detects an explicit creation/development request, it routes to your project's existing skills (writing, game dev, web, debugging…) per skills/daozhu/workflows/routing-workflow.md. Routed skills become 道樞's own modules — handed off with perception context, their internal flow untouched. If no matching skill exists, the cabinet responds instead.

Memory Layout (local_memory/)

local_memory/
├── daily/          # expenses, health, reminders, shopping, moods, notes, decisions
└── daozang/        # per-school strategy vault

Configuration (config.yaml, optional)

Copy mcp/config.yaml.example to mcp/config.yaml to override:

• timezone — display timezone
• currency — expense currency symbol (TWD/HKD/USD/JPY/CNY/EUR/GBP/KRW)
• review_weekday / review_time — weekly-report schedule
• energy_analysis_days — days needed before energy analysis is produced
• high_speed_threshold — concise-mode trigger threshold (words/min, used by the dispatch layer)

yaml is an optional dependency — without it, defaults are used silently.

Local dashboard (optional)

A loopback-only web UI for the same local_memory/ the MCP uses. Not a chat window.

pip install -r mcp/requirements-web.txt
# optional: point at an existing store (e.g. a Claude Code project copy)
# Windows:
set DAOZHU_MEMORY_DIR=C:\path\to\local_memory
# POSIX:
export DAOZHU_MEMORY_DIR=/path/to/local_memory
python -m web

Open http://127.0.0.1:8765. Three pages: board /, cabinet /cabinet, mirror /xinjing. Binds to 127.0.0.1 only. A left rail switches pages and themes (Daoist default / Confucian / Legalist / Strategist / Taiji) and remembers the choice in localStorage; below 720px the rail collapses to one-character icons. The board has a chat line that first matches rules (午餐吃了 150); optional DAOZHU_LLM_API_KEY classifies leftovers. Cabinet can convene a five-stage meeting (brief / deep / flash) and follow up one minister. The mirror still plays JSON; set DAOZHU_TIANJI_DIR to a local tianji-mcp checkout for real tarot / gua / fengshui / chart / bazi / ziwei / meihua. Optional thin desktop: python -m web.desktop. Optional native window (needs pywebview): python -m web.desktop --window. Optional VS Code sidebar lives in extension/ (local vsix only, not Marketplace).

Development

cd mcp
pip install -r requirements.txt
python -m pytest tests/          # 42 tests across 4 suites
python server.py                 # run MCP server over stdio

Contributing

See CONTRIBUTING.md and CHANGELOG.md.

License

Apache License 2.0 — see NOTICE for the project authors.