b6a478e1df
## Summary - Add portable installation tool (`install.sh`) for cross-machine setup - Add Claude.ai export files with proper YAML frontmatter - Add multi-agent-guide v2.0 with consolidated framework template - Rename `00-claude-code-setting` → `00-our-settings-audit` (avoid reserved word) - Add YAML frontmatter to 25+ SKILL.md files for Claude Desktop compatibility ## Commits Included - `93f604a` feat: Add portable installation tool for cross-machine setup - `9b84104` feat: Add Claude.ai export for portable skill installation - `f7ab973` fix: Add YAML frontmatter to Claude.ai export files - `3fed49a` feat(multi-agent-guide): Add v2.0 with consolidated framework - `3be26ef` refactor: Rename settings-audit skill and add YAML frontmatter Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2.3 KiB
2.3 KiB
MCP Server Token Estimates
Approximate token usage for common MCP servers when loaded.
Official Anthropic Servers
| Server | Tools | Tokens (est.) | Recommendation |
|---|---|---|---|
| Playwright | 20+ | 13,500 | Keep if testing |
| Filesystem | 8 | 4,000 | Essential |
| Memory | 5 | 3,000 | Optional |
| Brave Search | 2 | 1,500 | Alternative to web |
| Fetch | 2 | 1,200 | Keep for web access |
Third-Party Servers
| Server | Tools | Tokens (est.) | Recommendation |
|---|---|---|---|
| Notion | 10 | 5,000 | Keep if using Notion |
| GitHub | 25+ | 18,000 | Lazy load only |
| PostgreSQL | 15 | 8,000 | Lazy load only |
| Slack | 12 | 6,000 | Lazy load only |
| Figma | 15 | 10,000 | Lazy load only |
| Zapier | 50+ | 25,000+ | Avoid |
| Firecrawl | 5 | 3,500 | Useful for SEO |
| Perplexity | 2 | 1,500 | Alternative to web |
Token Budget Guidelines
Conservative Setup (High Available Context)
2-3 essential MCP servers
Total: ~10,000-15,000 tokens
Available: 185,000+ tokens (92%+)
Standard Setup
4-5 MCP servers
Total: ~20,000-30,000 tokens
Available: 170,000+ tokens (85%+)
Heavy Setup (Risk Zone)
6+ MCP servers
Total: 40,000+ tokens
Available: <160,000 tokens (<80%)
⚠️ Higher chance of "Exceed response limit"
Reduction Strategies
If using GitHub MCP heavily:
- Start conversation specifically for GitHub work
- Don't load other heavy servers simultaneously
If using multiple API servers:
- Choose ONE per category:
- Web search: Brave OR Perplexity (not both)
- Automation: Zapier OR n8n (not both)
- Database: PostgreSQL OR MySQL (not both)
If testing/development:
- Playwright OR Puppeteer (not both)
- Load only when actively testing
Impact of serverInstructions
Without serverInstructions:
- Full tool schemas loaded (~500 tokens per tool)
- Discovery overhead on each interaction
With serverInstructions:
- Reduced discovery queries
- More efficient tool selection
- Saves ~30-50% on tool-related tokens
Quick Reference
Token estimation formula:
Server tokens ≈ (number of tools × 400) + 500 base
Example:
- 10 tools = (10 × 400) + 500 = 4,500 tokens
- 25 tools = (25 × 400) + 500 = 10,500 tokens