AgentSkillsCN

mcp-guide

针对何时应使用 MCP 而非直接调用 API 提供指导

SKILL.md
--- frontmatter
name: mcp-guide
description: Guidance for when to use MCP vs direct API calls
version: 1.0.0
triggers:
  - mcp
  - which tool
  - how to call

MCP vs Skills/Direct API Guide

Decision framework for choosing between MCP tools and direct API calls.

Quick Decision

code
Need to call an external service?
├─ Is there a skill for it? → Use the skill
├─ Is cost control important? → Use direct API
├─ Need complex multi-step workflow? → Consider MCP
└─ Simple one-off call? → Direct API is fine

When to Use Skills (Direct API)

Prefer skills/direct API when:

ScenarioWhy
Cost-sensitive servicesControl model selection (Perplexity)
Simple queriesLess overhead than MCP
High-frequency callsFaster, no tool discovery
Need predictable behaviorSkills have explicit instructions

Services with skills:

ServiceSkillWhy not MCP
PerplexityresearchMCP defaults to expensive models
PostHoganalyticsDirect API is simpler
EmailemailHimalaya CLI, no MCP needed
GitHubgithubgh CLI is more reliable

When to Use MCP

Use MCP when:

ScenarioWhy
Complex multi-step workflowsMCP handles tool chaining
Need tool discoveryDon't know exact endpoint
Official MCP server existsBetter maintained
Service has many endpointsMCP abstracts complexity

Good MCP use cases:

ServiceMCP ServerUse Case
ApifyapifyWeb scraping (many actors)
DataForSEOdfs-mcpSEO data (many endpoints)
ProspectprospectLead enrichment
Google Workspacegoogle-mcpCalendar, Sheets, Drive

MCP Gotchas

1. Bad Defaults

Some MCP servers have expensive defaults:

MCPIssueSolution
Perplexity MCPDefaults to sonar-deep-researchUse research skill instead

2. Token Overhead

MCP tool calls include:

  • Tool discovery
  • Schema validation
  • Extra context

For simple calls, this overhead can exceed the actual API call.

3. Startup Time

MCP servers (especially npx-based) have cold start delays:

  • First call: 5-15 seconds
  • Subsequent: Normal

Decision Matrix

FactorUse SkillUse MCP
One endpoint
Many endpoints
Cost control needed
Complex workflow
Speed critical
Tool discovery needed
Explicit instructions exist

Available Skills

Check skills/ directory or use /list-skills for current skills.

SkillPurpose
researchPerplexity web search
analyticsPostHog queries
emailEmail via Himalaya
githubGitHub via gh CLI
lead-handlerProcess incoming leads

Available MCP Servers

Check workspace/mcporter.json for configured MCP servers.

Current servers:

  • apify - Web scraping actors
  • dfs-mcp - DataForSEO
  • prospect - Lead enrichment
  • google-mcp - Google Workspace

Examples

Bad: Using MCP for simple Perplexity search

code
❌ mcp__perplexity__search("latest AI news")
   → May use sonar-deep-research ($0.40+)

Good: Using research skill

code
✓ Use research skill with sonar model
   → Uses sonar ($0.006)

Good: Using MCP for Apify web scraping

code
✓ mcp__apify__call-actor for complex scraping
   → Many actors, MCP handles discovery

Adding New Integrations

  1. Simple service, one endpoint? → Create a skill
  2. Complex service, many endpoints? → Add MCP server to mcporter.json
  3. Cost-sensitive? → Always create a skill with explicit cost controls