tools-only MCP Tools — Prefer Over Manual Alternatives
MCP server tools are available as direct tool calls — just like Read, Grep, or Bash. They are NOT CLI commands. Do NOT run them via claude mcp call or any shell command. Call them directly as tools.
CORRECT — call the tool directly:
Tool: mcp__playwright__browser_snapshot
Parameters: {}
WRONG — do NOT shell out:
Bash: claude mcp call playwright browser_snapshot ... # WRONG — not a CLI command
MCP Server Overview
| Server | Prefix | Purpose | When to Use Instead of... |
|---|---|---|---|
| Playwright | mcp__playwright__ |
Live browser interaction, form filling, screenshots | Manual browser testing |
| Context7 | mcp__context7__ |
Library/framework documentation | Googling docs, outdated training data |
| Tavily | mcp__tavily__ |
Web search, URL extraction, site crawling | WebSearch, WebFetch for complex queries |
| Sequential Thinking | mcp__sequential-thinking__ |
Structured multi-step reasoning | Unstructured debugging of complex problems |
| Draw.io | mcp__drawio__ |
Create/edit diagrams with browser preview | Hand-crafting SVG, manual diagram tools |
Playwright MCP
For live browser interaction — navigating pages, clicking buttons, filling forms, taking screenshots.
| Key Rule | Why |
|---|---|
Always browser_snapshot before interacting |
Gives ref values needed for click/type/hover |
| Prefer snapshot over screenshot | Structured accessibility tree vs opaque image |
Use browser_wait_for for async pages |
Pages may not have rendered yet after navigation |
Do NOT use for writing Playwright test code — use the playwright-e2e skill instead. This is for live browser control.
Context7 MCP
For querying up-to-date library and framework documentation.
| Key Rule | Why |
|---|---|
Always resolve-library-id first |
Must resolve library name to /org/project ID before querying |
| Max 3 calls per tool per question | Avoid excessive API usage |
| Be specific in queries | "Next.js 15 middleware auth" beats "middleware" |
Common Library IDs (skip resolve for these)
| Library | ID |
|---|---|
| Next.js | /vercel/next.js |
| React | /facebook/react |
| Supabase JS | /supabase/supabase-js |
| TanStack Query | /tanstack/query |
| Zod | /colinhacks/zod |
Do NOT use for general web search — use Tavily.
Tavily MCP
For web search, content extraction, site crawling, and multi-source research.
| Key Rule | Why |
|---|---|
Start with tavily_search |
Low cost, fast — sufficient for most questions |
Use tavily_research sparingly |
Very expensive — multi-source synthesis |
| Include the year in queries | "topic 2026" gets current results |
tavily_map before tavily_crawl |
Understand site structure before extracting content |
Do NOT use for library docs — use Context7.
Sequential Thinking MCP
For structured multi-step reasoning on complex problems. Single tool: mcp__sequential-thinking__sequentialthinking.
| Key Rule | Why |
|---|---|
| Only use for genuinely complex problems | Simple problems don't need structured reasoning |
Use isRevision: true to correct earlier thoughts |
Maintains coherent reasoning chain |
Use branchFromThought for alternatives |
Explore "what if?" without losing main thread |
Start with totalThoughts: 5 |
Adjust with needsMoreThoughts: true if needed |
When to Use
- Bug with 3+ possible causes needing elimination
- Architectural decision with competing trade-offs
- Root cause analysis requiring hypothesis testing
When NOT to Use
- Simple errors with clear stack traces
- Straightforward implementation decisions
- Problems with known solutions
Parallel MCP Calls
Claude defaults to sequential MCP tool calls even when tools are independent. For multi-tool workflows, explicitly mark independent calls for parallel execution:
**Run these in parallel** (they are independent):
- `mcp__context7__searchDocs` → library docs
- `mcp__tavily__tavily_search` → web results
Independent MCP calls (different servers, no data dependencies) should always run in parallel to reduce latency.
When NOT to Use MCP Tools
- Writing SQL/migrations — use the
postgres-expertskill instead - Building application code — use
service-builder,server-action-builder,react-form-builder - Reading a specific known file —
Readis faster when you already know the exact path - Broad codebase search —
Grep/Globare more flexible for arbitrary pattern matching - Writing Playwright test code — use the
playwright-e2eskill (not the Playwright MCP server)