# 1. 克隆项目
git clone https://github.com/hanjiang-215/cognitive-exoskeleton-mcp.git
cd cognitive-exoskeleton-mcp

# 2. 安装依赖并构建
npm install
npm run build

# 3. 启动（默认零配置）
node dist/index.js

{
  "mcpServers": {
    "cognitive-exoskeleton": {
      "command": "node",
      "args": ["<项目路径>/dist/index.js"],
      "env": {
        "LLM_MODE": "sampling"
      }
    }
  }
}

codebuddy mcp add cognitive-exoskeleton \
  --command "node" \
  --arg "<项目路径>/dist/index.js" \
  --env LLM_MODE=sampling

export LLM_MODE=direct
export LLM_API_BASE="https://api.openai.com/v1"
export LLM_API_KEY="sk-你的密钥"
export LLM_MODEL_NAME="gpt-4o-mini"
node dist/index.js

$env:LLM_MODE = "direct"
$env:LLM_API_BASE = "https://api.openai.com/v1"
$env:LLM_API_KEY = "sk-你的密钥"
$env:LLM_MODEL_NAME = "gpt-4o-mini"
node dist/index.js

set LLM_MODE=direct
set LLM_API_BASE=https://api.openai.com/v1
set LLM_API_KEY=sk-你的密钥
set LLM_MODEL_NAME=gpt-4o-mini
node dist/index.js

{
  "mcpServers": {
    "cognitive-exoskeleton": {
      "command": "node",
      "args": ["<项目路径>/dist/index.js"],
      "env": {
        "LLM_MODE": "direct",
        "LLM_API_BASE": "https://api.openai.com/v1",
        "LLM_API_KEY": "sk-你的密钥",
        "LLM_MODEL_NAME": "gpt-4o-mini"
      }
    }
  }
}

用户：请把这篇笔记导入知识图谱：
"分布式系统遵循 CAP 定理，真正选择是在 CP 和 AP 之间。"
→ 自动抽取 CAP定理、一致性、可用性等实体及关系

用户：我对 CAP 定理了解多少？
→ 从图谱检索相关实体，LLM 推理后返回结构化答案

用户：我正在写关于数据库一致性模型的文字...
→ 召回 3 个月前关于 CAP 定理的旧笔记

用户：分布式系统和机器学习之间有什么隐藏联系？
→ "你的'共识算法'和'反向传播'可能有关联：都通过迭代反馈达成全局一致性"

用户：分析我对"神经网络"理解的盲点
→ "你了解 CNN、RNN、Transformer，但缺少：图神经网络、神经架构搜索、模型压缩..."

用户：展示我的知识图谱整体结构
→ 3 个孤岛、桥梁概念"一致性"、稀疏区域：系统安全和性能优化

用户：碰撞"分布式系统"和"神经科学"
→ "大脑的神经可塑性类似于分布式系统的自适应拓扑。突触修剪 ≈ 节点退役。"

MCP 客户端 (Cursor / CodeBuddy / Cline)
        │ stdio (JSON-RPC)
        │ + sampling/createMessage (Sampling 模式)
        ▼
┌──────────────────────────────────────┐
│  Cognitive Exoskeleton MCP Server   │
│                                      │
│  8 个 MCP 工具                       │
│         │                            │
│  知识图谱引擎 (SQLite + 图算法)       │
│         │                            │
│  LLM 双模式: Sampling / Direct       │
└──────────────────────────────────────┘

nodes (id, type, name, summary, domain, aliases, source_file,
       first_seen_at, last_seen_at, mention_count)
edges (id, source_id, target_id, relation, confidence, evidence, created_at)
notes_index (file_path, content_hash, node_ids, last_ingested_at)
evolution_log (id, node_id, snapshot_at, belief_summary, trigger_note, source_file)
topology_cache (snapshot_at, isolated_clusters, bridge_nodes, density_map, summary)
serendipity_log (id, node_a, node_b, hypothesis, user_feedback, created_at)

git clone https://github.com/hanjiang-215/cognitive-exoskeleton-mcp.git
cd cognitive-exoskeleton-mcp
npm install
npm run build

# Zero-config — automatically reuses your MCP client's LLM via Sampling
node dist/index.js

{
  "mcpServers": {
    "cognitive-exoskeleton": {
      "command": "node",
      "args": ["<project-path>/dist/index.js"],
      "env": {
        "LLM_MODE": "sampling"
      }
    }
  }
}

codebuddy mcp add cognitive-exoskeleton \
  --command "node" \
  --arg "<project-path>/dist/index.js" \
  --env LLM_MODE=sampling

export LLM_MODE=direct
export LLM_API_BASE="https://api.openai.com/v1"
export LLM_API_KEY="sk-..."
export LLM_MODEL_NAME="gpt-4o-mini"
node dist/index.js

$env:LLM_MODE = "direct"
$env:LLM_API_BASE = "https://api.openai.com/v1"
$env:LLM_API_KEY = "sk-..."
$env:LLM_MODEL_NAME = "gpt-4o-mini"
node dist/index.js

set LLM_MODE=direct
set LLM_API_BASE=https://api.openai.com/v1
set LLM_API_KEY=sk-...
set LLM_MODEL_NAME=gpt-4o-mini
node dist/index.js

User: Ingest this note: "Distributed systems follow the CAP theorem..."
→ Extracts CAP Theorem, Consistency, Availability, etc. + relationships

User: What do I know about the CAP theorem?
→ Retrieves related entities, LLM reasons and returns structured answer

User: I'm writing about database consistency models...
→ Recalls notes from 3 months ago about CAP theorem

User: Hidden connections between distributed systems and ML?
→ "Your 'consensus algorithms' and 'backpropagation' may be related:
   both achieve global consistency through iterative feedback"

User: Blindspots in my understanding of neural networks?
→ "You know CNNs, RNNs, Transformers, but missing: GNNs, NAS, model compression..."

User: Show me the overall structure of my knowledge graph
→ 3 islands, bridge concept "consistency", sparse: security, optimization

User: Spark between distributed-systems and neuroscience
→ "Neural plasticity ≈ adaptive topology. Synaptic pruning ≈ node decommissioning."

MCP Client (Cursor / CodeBuddy / Cline)
        │ stdio (JSON-RPC)
        │ + sampling/createMessage (Sampling mode)
        ▼
┌──────────────────────────────────────┐
│  Cognitive Exoskeleton MCP Server   │
│                                      │
│  8 MCP Tools                         │
│         │                            │
│  Knowledge Graph Engine              │
│  (SQLite + graph algorithms)         │
│         │                            │
│  LLM Client                          │
│  (Sampling / Direct dual mode)       │
└──────────────────────────────────────┘

nodes (id, type, name, summary, domain, aliases, source_file,
       first_seen_at, last_seen_at, mention_count)
edges (id, source_id, target_id, relation, confidence, evidence, created_at)
notes_index (file_path, content_hash, node_ids, last_ingested_at)
evolution_log (id, node_id, snapshot_at, belief_summary, trigger_note, source_file)
topology_cache (snapshot_at, isolated_clusters, bridge_nodes, density_map, summary)
serendipity_log (id, node_a, node_b, hypothesis, user_feedback, created_at)

npm install          # Install dependencies
npm run dev          # Watch mode (auto-rebuild)
npm run build        # Production build
node dist/index.js   # Start server

npm run build

# In your MCP client:
# 1. ingest_note → "examples/sample-notes/distributed-systems.md"
# 2. ingest_note → "examples/sample-notes/neural-networks.md"
# 3. query_mind → "What do I know about consensus?"
# 4. detect_blindspots → topic = "distributed systems"
# 5. analyze_cognitive_topology → (no arguments)
# 6. discover_connections → topic = "consensus"
# 7. spark_serendipity → domain_a = "distributed-systems", domain_b = "machine-learning"