872401be72
New resource note with 6 core methodologies, community best practices, pitfalls, and practical examples. Three zettelkasten notes extract key insights: hook vs prompt reliability, MCP context tradeoffs, and the instinct learning system. Updated existing guides with cross-links.
29 lines
899 B
Markdown
29 lines
899 B
Markdown
---
|
||
created: "2026-03-19 12:02"
|
||
type: zettel
|
||
tags: [claude-code, context-window, performance, mcp]
|
||
source: "https://github.com/affaan-m/everything-claude-code"
|
||
---
|
||
|
||
# MCP 数量与上下文窗口的反比关系
|
||
|
||
每个启用的 MCP Server 都会占用上下文窗口空间（工具定义、schema 描述等）。实测数据：
|
||
|
||
- 0-5 个 MCP: 有效上下文接近 200K tokens
|
||
- 10+ 个 MCP: 有效上下文降至 ~70K tokens（降幅 65%）
|
||
|
||
**最佳实践**:
|
||
- 活跃 MCP <= 10 个，活跃工具总数 <= 80 个
|
||
- 用 `disabledMcpServers` 动态禁用不用的
|
||
- 重操作优先用 CLI（如 `gh`）而不是 MCP
|
||
- 使用 `llms.txt` 模式获取文档，避免 MCP 常驻开销
|
||
|
||
这是 Claude Code 性能优化中**最重要的单一因素** — 比任何 Agent 或 Skill 都重要。
|
||
|
||
---
|
||
|
||
## Related
|
||
|
||
- [[Everything Claude Code 方法论与最佳实践]]
|
||
- [[Hook驱动优于提示词驱动]]
|