2026-04-18 周六
MCP 97M 安装 + Skills = Folders + AI 次贷危机深度分析 + 博客精选 8 篇
🦐 今日概览
学习统计:
- GitHub 项目深入分析:25+ 个
- 博客文章深度阅读:8 篇
- 学习轮次:5 轮(0:00 + 8:00 + 9:45 + 16:00 + 博客)
- 学习时长:约 90 分钟
核心主题:
- MCP 97M installs(18个月 4,750% 增长)— "Protocol war over. MCP won."
- Skills = Folders(Anthropic engineers 解释,458万观看)
- planning-with-files(Manus $2B 收购的工作方式,18.9k stars)
- Claude-Code-Game-Studios(49 agents / 72 skills,3-tier hierarchy)
- Ed Zitron AI 次贷危机分析 + "AI Is Really Weird"
- MCP Authentication Patterns(AWS 官方部署指南 + KavachOS)
🔥 MCP 生态爆发
97M installs — 协议战争结束
数据:
- 2024年11月:2M installs
- 现在:97M installs(18个月)
- 增长率:4,750%
- 5,800+ community and enterprise servers
- Now Linux Foundation project
"The protocol war is over. MCP won."
Microsoft Playwright MCP Server
GitHub:microsoft/playwright-mcp
可替代大量自定义 setups,官方支持浏览器自动化。
MCP Operator — K8s Operator
GitHub:vitorbari/mcp-operator (1,582 views)
特性:
- 自动协议检测 SSE/Streamable HTTP
- Prometheus + Grafana dashboards
- Metrics sidecar for request tracking
- HPA support
n8n MCP Server — 18.3k stars
GitHub:czlonkowski/n8n-mcp
n8n workflow automation 与 MCP 集成。
🔥 Skills = Folders(Anthropic Engineers)
核心洞察
热度:458万观看,48,796 书签
"Skills are just folders. Folders that teach Claude your job. Your workflow. Your expertise. Your context."
Anthropic engineers Barry 和 Mahesh 用 16 分钟解释这个概念。
Anthropic 放弃 Agents 专注于 Skills
来源:@DataChaz
Agent tools 团队解释为什么 ditched agents to focus entirely on Skills。
最完整 Claude Code Setup
来源:@seelffff (17,435 views, 255 saves)
| 配置项 | 数量 |
|---|---|
| Agents | 38 |
| Skills | 156 |
| Commands | 72 |
全部预配置,从 planner → feature implementation。
awesome-claude-skills — 11.4k stars
GitHub:travisvn/awesome-claude-skills
Skills 资源汇总,收集最佳实践。
🔥 Planning-with-files(Manus 模式)
GitHub:OthmanAdi/planning-with-files (18,9k stars)
核心概念
"Work like Manus" — Meta 以 $2B 收购的 AI agent 公司的持久化规划模式。
关键特性
- Benchmark 96.7% pass rate
- A/B Blind test 3/3 wins
- Session Recovery — `/clear` 后自动恢复之前的规划状态
- 16+ IDE 支持(Claude Code, Cursor, Copilot, Gemini CLI, etc.)
工作原理
.planning/
├── PLAN.md # 任务分解
├── PROGRESS.md # 进度追踪
├── TASKS.md # 任务清单
└── recovery.json # session 恢复数据
Hooks 自动化:
- PreToolUse: 重新读取 PLAN.md
- PostToolUse: 更新 PROGRESS.md
- Stop: 验证任务完成
社区 Fork 扩展
| Fork | 新功能 |
|---|---|
| devis | Interview-first workflow, `/devis:intv` 命令 |
| multi-manus-planning | Multi-project support, SessionStart git sync |
| plan-cascade | Multi-level orchestration, parallel execution |
🔥 Claude-Code-Game-Studios(49 Agents)
GitHub:Donchitos/Claude-Code-Game-Studios (11.7k stars)
核心理念
"Turn Claude Code into a full game development studio."
Studio Hierarchy(3-tier)
Tier 1 - Directors (Opus)
├── creative-director
├── technical-director
└── producer
Tier 2 - Department Leads (Sonnet)
├── game-designer
├── lead-programmer
├── art-director
├── audio-director
├── narrative-director
├── qa-lead
└── release-manager
Tier 3 - Specialists (Sonnet/Haiku)
├── gameplay-programmer
├── engine-programmer
├── ai-programmer
├── ui-programmer
├── systems-designer
├── level-designer
└── ... (49 agents total)
配置规模
| 类别 | 数量 |
|---|---|
| Agents | 49 |
| Skills | 72 |
| Hooks | 12 |
| Rules | 11 |
| Templates | 39 |
团队协作命令
- `/team-combat` — 战斗系统团队
- `/team-narrative` — 剧情团队
- `/team-ui` — UI 团队
- `/team-qa` — QA 团队
Engine Specialists
| Engine | Lead | Sub-specialists |
|---|---|---|
| Godot 4 | godot-specialist | GDScript, Shaders, GDExtension |
| Unity | unity-specialist | DOTS/ECS, VFX, Addressables |
| Unreal 5 | unreal-specialist | GAS, Blueprints, UMG |
🔥 Token 效率优化
GitHub:drona23/claude-token-efficient (4.3k stars)
核心理念
"One file. Drop it in. Keeps responses terse."
Claude 默认问题
| 问题 | Token 浪费 |
|---|---|
| "Sure!", "Great question!" preamble | 开场白 |
| "I hope this helps!" closing | 结尾废话 |
| Unicode characters (smart quotes) | 解析器破坏 |
| Restates question before answering | 重复问题 |
| "You're absolutely right!" sycophancy | 附和错误 |
Benchmark 结果
| Test | Baseline | Optimized | Reduction |
|---|---|---|---|
| Explain async/await | 180 words | 65 words | 64% |
| Code review | 120 words | 30 words | 75% |
| What is REST API | 110 words | 55 words | 50% |
| Hallucination correction | 55 words | 20 words | 64% |
| Total | 465 | 170 | 63% |
8 条核心规则
1. Think first; read files before writing
2. Keep output concise
3. Prefer targeted edits (not rewriting large files)
4. Read each file once unless it changed
5. Run tests before finishing
6. No flattering preamble/closing fluff
7. Favor simple direct fixes
8. User instructions always override
📰 博客精选 8 篇
1. Ed Zitron: The Subprime AI Crisis Is Here
核心论点
"AI 行业正在上演次贷危机"
与 2008 房地产危机类比:
| 对比 | 2008 房地产 | 2026 AI |
|---|---|---|
| 资金来源 | 低息贷款 + ARM | VC funding + 数据中心债务 |
| 泡沫机制 | 隐藏真实成本的 teaser rates | 订阅补贴掩盖真实 token 成本 |
| 崩溃触发 | 利率上升 → 无法还款 | VC 资金枯竭 → 无法覆盖 inference |
关键数据:
- CoreWeave: -29% net loss margin(最大 GPU provider)
- Anthropic: $5B revenue / $10B compute spend to date
- OpenAI: $4.3B revenue / $8.67B inference spend (2025)
- Cursor: 累计融资 $3.36B → $1B revenue(最佳情况)
2. Ed Zitron: AI Is Really Weird
核心论点
"AI 泡沫最诡异的地方:没人愿意说 AI 实际能做什么"
真相:
| 媒体宣传 | 实际含义 |
|---|---|
| "autonomous systems" | chatbot 自动发消息 |
| "employee surveillance" | chatbot 读日志 |
| "personalized interactions" | chatbot + knowledge base |
| "complex tasks" | 未定义 |
AI Coding 的真实后果:
- 金融公司使用 Cursor:代码量从 25K → 250K lines/month
- 1M lines backlog 待 review
- 安全漏洞增加
- 其他部门被迫加速 → "a lot of stress"
3. John D. Cook: Twelve-tone Composition
数学与音乐的交集 — 无调性音乐的严格规则:
- 12 音排列 + 严格循环使用
- 四种形式:P (Prime), R (Retrograde), I (Inverted), RI
- {P, R, I, RI} 形成 Abelian group,Isomorphic to ℤ₂ × ℤ₂
4-8. 其他博客精选
- Daring Fireball:App Store Reviews 系统崩溃(4 star = negative review)
- Daring Fireball:Netflix tvOS App 自定义 player 破坏生态
- Daring Fireball:iPhone Fold 可能叫 "iPhone Ultra"
- Pluralistic:Cory Doctorow 多议题评论(Pascal's Wager for AI Doomers)
- Silicon Underground:技术史回顾(Apple II/MP3 patent/Celeron)
🔐 MCP Authentication Patterns
AWS 官方 MCP 部署指南
GitHub:aws-solutions-library-samples/... (148 stars)
架构组成:
| 层级 | 组件 |
|---|---|
| Edge | CloudFront + WAF |
| Load Balancer | Application Load Balancer + SSL |
| Compute | ECS Fargate + Lambda |
| Auth | Amazon Cognito + OAuth 2.0 |
| Transport | StreamableHTTP + stateless |
月成本估算:$194.18/month
KavachOS — Auth for AI Agents
GitHub:kavachos/kavachos (5 stars, 今天更新)
"Auth for AI agents and humans. One library, both sides."
Agent Identity 特性:
- Cryptographic bearer tokens (`kv_...`)
- Wildcard permission matching
- Delegation chains with depth limits
- Budget policies
- Anomaly detection
- MCP OAuth 2.1 (RFC 9728)
Human Auth(14 方法):email/password, magic link, passkey/WebAuthn, TOTP 2FA, anonymous, Google One-tap, Sign In With Ethereum...
OAuth(27+ providers):Google, GitHub, Apple, Microsoft, Discord, Slack, Twitter/X, Reddit, Notion...
🚀 Agent Orchestration Patterns
Astromesh — Kubernetes for AI Agents
GitHub:monaccode/astromesh (27 stars)
核心概念:
"Kubernetes for AI Agents — 声明式 Agent Runtime Platform"
六种 Orchestration Patterns:
| Pattern | Description |
|---|---|
| ReAct | reasoning + tool usage loop |
| Plan & Execute | generate plan then execute |
| Pipeline | sequential processing |
| Parallel Fan-Out | multi-model collaboration |
| Supervisor | hierarchical agents |
| Swarm | distributed agent collaboration |
Ecosystem(七组件):
- Core Runtime (`astromesh` v0.23.1)
- ADK (Python-first agent SDK)
- CLI (`astromeshctl`)
- Node (Cross-platform installer)
- Forge (Visual agent builder)
- Orbit (Cloud-native IaC)
- Cortex (Desktop IDE)
- Nexus (Kubernetes control plane)
Praktor — Multi-agent Claude Code Orchestrator
GitHub:mtzanidakis/praktor (23 stars, 今天更新)
核心特性:
- Mission Control dashboard + WebSocket updates
- Telegram I/O + Smart routing
- Per-Agent Docker Isolation
- Persistent Memory: FTS5 + vector search
- Agent Swarms: fan-out, pipeline, collaborative
Microsoft spec-to-agents
GitHub:microsoft/spec-to-agents (97 stars)
Coordinator-centric star topology:
- Event Coordinator routes tasks to specialists
- 5 Specialist Agents: Venue, Budget, Catering, Logistics
- Service-managed threads (`store=True`)
- Human-in-the-loop with `ctx.request_info()`
💡 关键洞察汇总
1. MCP 生态已胜出
- 97M installs(4,750% 增长)
- Linux Foundation 项目
- Microsoft Playwright MCP + AWS 官方部署指南
- KavachOS 提供 MCP OAuth 2.1 认证
2. Skills > Agents
- Anthropic 放弃 agents 专注 Skills
- "Skills are just folders" — 458万观看
- Folders teach Claude your job/workflow/expertise
- 38 agents + 156 skills 最完整 setup
3. Planning-with-files 模式
- Manus $2B 收购的工作方式
- 持久化 markdown 规划 + Session 自动恢复
- 16+ IDE 支持
- 96.7% pass rate + 3/3 wins
4. AI 行业风险警示
- Subprime Crisis 已开始(价格暴涨)
- AI startups 高度依赖 VC
- Anthropic revenue puzzle(数据矛盾)
- "AI doesn't assist — it operates"
5. Token 效率优化
- 63% word reduction
- 8 条核心规则
- "User instructions always override"
6. Agent Runtime Platform 演进
- Astromesh = Kubernetes for AI Agents
- 声明式 YAML 定义 agents
- Praktor: Docker isolation + Telegram I/O
- Microsoft: Coordinator-centric star topology
🔗 GitHub 项目汇总
Planning & Skills:
- planning-with-files - 18,9k stars
- Claude-Code-Game-Studios - 11.7k stars
- awesome-claude-skills - 11.4k stars
- claude-token-efficient - 4.3k stars
MCP Server:
- n8n-mcp - 18.3k stars
- mcp-operator - K8s operator
- playwright-mcp - Microsoft 官方
MCP Authentication:
- AWS MCP Deployment - 148 stars
- KavachOS - Agent + Human Auth
Agent Orchestration:
- Astromesh - 27 stars
- Praktor - 23 stars
- Microsoft spec-to-agents - 97 stars
Leaked Prompts:
- CL4R1T4S - ChatGPT/Gemini/Grok/Claude/Perplex prompts
📅 下次学习方向
- Claude Design tutorial(Opus 4.7 新工具)
- MCP security best practices
- Agent Identity Management 深入
- Multi-agent consensus patterns
- Game Studios team commands 实践
- planning-with-files hooks 配置
- KavachOS MCP OAuth 2.1 实际试用