MCP Economy Skill

Purpose

Reduce token burn by ensuring MCP tools are only used when local context is insufficient. This skill provides explicit guidance on when to use external tools and when to avoid them.

Why This Matters

MCP tools add tokens to context quickly. Some servers (like GitHub or broad web search) can add thousands of tokens per call. This skill helps you:

• •Preserve context window for actual work
• •Reduce API costs
• •Avoid unnecessary external calls
• •Make deliberate, cost-aware decisions

Core Rules

Rule 1: Default to Local

Always try local reasoning and available files first. Most tasks can be solved with:

• •Files already in context
• •Previous conversation history
• •Built-in knowledge

Rule 2: MCPs Only When Needed

Use MCPs only when local context is missing or ambiguous:

• •Missing documentation for a library
• •Need to find code patterns across GitHub
• •User explicitly asks for web/external research

Rule 3: Prefer Lightweight MCPs

Use this priority order:

1. •context7 - Docs lookup (low cost, high utility)
2. •grep_app - GitHub code search (medium cost, targeted)
3. •zai-zread - Repo structure/docs (medium cost)
4. •exa/zai-search - Web search (higher cost, use sparingly)

Rule 4: Cap Expensive Queries

• •exa/web search: 1-2 queries per task unless user explicitly asks for more
• •GitHub search: Narrow scope first, broaden only if needed
• •Notion/Linear: Query only what's needed, avoid broad fetches

Rule 5: Summarize and Stop

When you use an MCP:

• •Extract only the relevant information
• •Summarize findings succinctly
• •Don't chain multiple MCP calls unless necessary

Decision Checklist

Before using an MCP, ask yourself:

[ ] Do I already have the answer locally?
[ ] Can I solve this with files already in context?
[ ] Is a small doc lookup (context7) enough?
[ ] Is GitHub code search specifically required?
[ ] Does the user explicitly ask for web discovery?
[ ] Have I already used my MCP budget for this task?

If you answer "yes" to the first two questions, do not use an MCP.

Token Budget Guidelines

Total external budget per task: ~5000-10000 tokens unless user explicitly approves more.

When to Use Each MCP

context7 (Docs Lookup)

Use when:

• •Need API documentation for a library
• •Looking up function signatures
• •Understanding library concepts

Avoid when:

• •You already know the API
• •The library is well-known (lodash, react, etc.)
• •Information is in project files

grep_app (GitHub Search)

Use when:

• •Looking for real-world usage patterns
• •Finding how others solved a specific problem
• •Researching open-source implementations

Avoid when:

• •The pattern is obvious
• •Local codebase has examples
• •User hasn't asked for external research

exa/zai-search (Web Search)

Use when:

• •User explicitly asks for web research
• •Documentation is missing or outdated
• •Need recent/current information

Avoid when:

• •context7 or grep_app can answer
• •The question is about code (use grep_app instead)
• •Local context is sufficient

linear/notion (Project Management)

Use when:

• •User asks about tickets/issues
• •Need to create/update project items
• •Looking up project documentation

Avoid when:

• •Just coding tasks
• •User hasn't mentioned project management

Anti-Patterns (Avoid These)

Token Burner Patterns

• •Using web search for every question
• •Chaining multiple MCP calls without summarizing
• •Fetching entire repos when you only need one file
• •Running broad searches before trying narrow ones

Lazy Patterns

• •Using MCP instead of thinking through the problem
• •Fetching docs for libraries you already know
• •Searching GitHub for trivial patterns

Overkill Patterns

• •Using exa when context7 would suffice
• •Running multiple searches for the same concept
• •Fetching full Notion databases when you need one page

Examples

Good: Minimal MCP Usage

User: "How do I use zod for form validation?"

Thought: I know zod well. Let me check if there's a specific pattern
the user wants, or if my knowledge is sufficient.

Action: Answer from knowledge first. If user needs specifics,
use context7 for zod docs (1 call max).

Good: Targeted Search

User: "Find how other projects handle rate limiting in Express"

Thought: This requires external research. Use grep_app with
a narrow query.

Action: grep_app search for "express rate limit middleware" (1 call)
Summarize top 2-3 patterns found.

Bad: Token Burning

User: "How do I use lodash?"

Bad Action: 
1. exa search for "lodash tutorial" 
2. context7 lookup for lodash
3. grep_app search for "lodash usage patterns"

This burns ~8000+ tokens for something you already know.

Quick Reference

Summary

1. •Local first - Most answers are already available
2. •Lightweight MCPs - context7 > grep_app > exa
3. •Budget conscious - Cap expensive calls
4. •Summarize - Extract what you need, don't chain
5. •Ask if unsure - User can approve more MCP usage

Default stance: Do not use MCPs unless clearly necessary.