connect-mcp-server
ronnycoding/.claude
本指南說明如何透過 HTTP、stdio 及 SSE 傳輸協定,將 MCP(模型上下文協定)伺服器連接到 Claude Code。內容涵蓋安裝、設定、驗證、環境變數及安全性。當使用者希望連線至 MCP 伺服器、新增整合功能、設定外部服務,或提及 MCP、伺服器、整合功能或外部工具時,請參閱本指南。
...展開全部關於connect-mcp-server
`connect-mcp-server` 技能可讓 MCP(模型上下文協定)伺服器與 Claude Code 無縫整合,提供一種標準化方法,將 AI 助理連接至外部服務與資料來源。此技能解決了從多個平台存取及管理工具、資源與提示詞的挑戰,且無需進行複雜的手動設定。 透過支援 HTTP、stdio 以及(已廢棄的)SSE 傳輸協定,使用者能高效連線至本地與遠端伺服器,從而簡化與 API、資料庫及生產力工具的互動流程。
此技能的主要特色包括靈活的傳輸選項,可透過 HTTP 連線至雲端服務、透過 stdio 連線至本地伺服器,以及透過 SSE 處理舊版應用情境。它提供指令來新增伺服器、設定認證標頭或環境變數,並測試連線以確保設定正確。 使用者可透過 Claude Code 並使用簡單的前綴,存取 MCP 提供的工具、資源和提示,從而流暢地執行各項任務,例如建立 GitHub 問題、管理 Notion 頁面,或與自訂 API 互動。 此技能支援 OAuth 2.0 驗證、本地系統整合,以及針對本地伺服器的離線運作,全面涵蓋各種開發與運作情境。
此技能非常適合需要將外部服務整合至 Claude Code 工作流程的開發人員、資料工程師及 AI 愛好者。它支援專案層級、使用者層級及本機範圍,以滿足不同的協作與部署需求。 應用案例包括自動化專案管理任務、從連線服務擷取及更新資料、整合專屬 API,以及運用本地系統工具來強化 AI 功能。透過簡化 MCP 伺服器連線,此技能讓使用者能夠擴展 Claude Code 的功能,並在多個平台上優化 AI 驅動的工作流程。
常見問題
對於遠端服務,我應該使用哪種傳輸方式?
對於遠端雲端服務,建議使用 HTTP 傳輸方式,因為它支援 JSON-RPC、可在不同網路間運作,並包含 OAuth 2.0 支援。
我可以在沒有網路連線的情況下,在本地端執行 MCP 伺服器嗎?
可以,stdio 傳輸方式可讓您運行具備直接系統存取權限及離線功能的本地伺服器。
如何對需要 API 金鑰或憑證的 MCP 伺服器進行驗證?
對於 HTTP 伺服器,您可以透過 HTTP 標頭提供憑證;對於 stdio 伺服器,則可透過環境變數提供憑證。您亦可透過 Claude Code 中的 `/mcp` 指令完成 OAuth 驗證。
SSE 傳輸方式是否仍受支援?
SSE 傳輸方式已不再推薦使用,建議新整合專案改用 HTTP 伺服器。
在 Windows 上使用本地伺服器時,是否需要特殊設定?
是的,在 Windows 上使用 npx 指令時,應將其以 `cmd /c` 包裹,並在指定指令前使用 `--` 作為分隔符。
Connect MCP Server Guide
This skill helps you connect MCP (Model Context Protocol) servers to Claude Code, enabling integrations with external services like GitHub, Notion, databases, project management tools, and custom APIs.
Quick Start
When connecting an MCP server, follow this workflow:
- Identify the server - Which service do you want to integrate?
- Choose transport - HTTP (remote), stdio (local), or SSE (deprecated)
- Gather credentials - API keys, tokens, or OAuth requirements
- Select scope - Local (default), project (.mcp.json), or user (cross-project)
- Add server - Use
claude mcp addcommand - Authenticate - Use
/mcpin Claude Code for OAuth - Test connection - Verify server status and available tools
- Use resources/prompts - Access via
@and/prefixes
What is MCP?
Model Context Protocol (MCP) is a standard for connecting AI assistants to external services and data sources. MCP servers provide:
- Tools: Functions Claude can execute (e.g., create GitHub issue)
- Resources: Data Claude can reference (e.g., @github:issue://123)
- Prompts: Pre-built commands (e.g., /mcp__github__review-pr)
Transport Protocols
1. HTTP Servers (Recommended)
Best for: Remote cloud-based servicesProtocol: HTTP/HTTPS with JSON-RPC
Basic syntax:
claude mcp add --transport http <name> <url>
Examples:
Notion integration:
claude mcp add --transport http notion https://mcp.notion.com/mcp
GitHub integration:
claude mcp add --transport http github https://mcp.github.com
Custom API with authentication:
claude mcp add --transport http secure-api https://api.example.com/mcp \ --header "Authorization: Bearer your-token-here"Multiple headers:
claude mcp add --transport http api https://api.example.com/mcp \ --header "Authorization: Bearer token" \ --header "X-API-Version: v2"
Advantages:
- No local installation required
- Works across networks
- Managed by service provider
- Automatic updates
- OAuth 2.0 support
2. Stdio Servers (Local)
Best for: Local tools requiring system accessProtocol: Standard input/output (stdin/stdout)
Basic syntax:
claude mcp add --transport stdio <name> <command> [args...]
Examples:
Using npx (recommended for Node.js servers):
claude mcp add --transport stdio filesystem -- npx -y @modelcontextprotocol/server-filesystem /path/to/allowed/directory
With environment variables:
claude mcp add --transport stdio airtable \ --env AIRTABLE_API_KEY=your-key-here \ -- npx -y airtable-mcp-server
Direct executable:
claude mcp add --transport stdio custom-server -- /usr/local/bin/my-server --arg1 value1
Windows-specific (requires cmd /c wrapper):
claude mcp add --transport stdio github-local -- cmd /c npx -y @modelcontextprotocol/server-github
Python server:
claude mcp add --transport stdio python-server -- python3 /path/to/server.py
Advantages:
- Direct file system access
- No network latency
- Full control over installation
- Works offline
- Can integrate system utilities
Important notes:
- The
--separator is required before the command - Use
-yflag with npx to auto-install - Windows requires
cmd /cwrapper for npx
3. SSE Servers (Deprecated)
Status: Deprecated - use HTTP servers instead
Syntax (if needed):
claude mcp add --transport sse <name> <url>
Note: SSE (Server-Sent Events) transport is being phased out. Migrate to HTTP servers where available.
Configuration Scopes
Local Scope (Default)
Storage: User settings (not shared)Use for: Personal project-specific servers
claude mcp add --transport http notion https://mcp.notion.com/mcp
No --scope flag means local scope.
Project Scope
Storage: .mcp.json in project root (shared via git)Use for: Team-wide integrations
claude mcp add --transport http github https://mcp.github.com --scope project
Creates: .mcp.json file:
{ "mcpServers": { "github": { "type": "http", "url": "https://mcp.github.com" } }}Important: Claude Code prompts for approval before using project-scoped servers from .mcp.json files for security.
User Scope
Storage: User settings (cross-project)Use for: Servers you use across all projects
claude mcp add --transport stdio filesystem \ --scope user \ -- npx -y @modelcontextprotocol/server-filesystem ~/DocumentsAvailable in all projects on your machine.
Manual Configuration (.mcp.json)
HTTP Server Configuration
{ "mcpServers": { "github": { "type": "http", "url": "https://mcp.github.com", "headers": { "Authorization": "Bearer ${GITHUB_TOKEN}", "X-Custom-Header": "value" } }, "notion": { "type": "http", "url": "https://mcp.notion.com/mcp" } }}Stdio Server Configuration
{ "mcpServers": { "filesystem": { "type": "stdio", "command": "npx", "args": [ "-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/directory" ] }, "postgres": { "type": "stdio", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres"], "env": { "POSTGRES_CONNECTION_STRING": "${DATABASE_URL}" } }, "custom-python": { "type": "stdio", "command": "python3", "args": ["/path/to/server.py", "--config", "config.json"], "env": { "PYTHONPATH": "/custom/python/path", "API_KEY": "${MY_API_KEY}" } } }}Complete Example with Multiple Servers
{ "mcpServers": { "github": { "type": "http", "url": "https://mcp.github.com", "headers": { "Authorization": "Bearer ${GITHUB_TOKEN}" } }, "filesystem": { "type": "stdio", "command": "npx", "args": [ "-y", "@modelcontextprotocol/server-filesystem", "${HOME}/projects" ] }, "postgres": { "type": "stdio", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres"], "env": { "POSTGRES_CONNECTION_STRING": "${DATABASE_URL:-postgresql://localhost:5432/mydb}" } }, "slack": { "type": "http", "url": "https://mcp.slack.com", "headers": { "Authorization": "Bearer ${SLACK_TOKEN}" } } }}Environment Variables
Variable Expansion
Claude Code supports environment variable expansion:
Syntax:
${VAR}- Expands to environment variable value${VAR:-default}- Uses default if variable not set
Supported locations:
- Command paths
- Arguments
- Environment settings
- HTTP headers
Examples:
{ "mcpServers": { "api": { "type": "http", "url": "https://api.example.com/mcp", "headers": { "Authorization": "Bearer ${API_TOKEN}", "X-Environment": "${ENVIRONMENT:-production}" } }, "local-server": { "type": "stdio", "command": "${HOME}/bin/my-server", "args": ["--config", "${CONFIG_PATH:-/etc/server.conf}"], "env": { "LOG_LEVEL": "${LOG_LEVEL:-info}", "DATA_DIR": "${DATA_DIR}" } } }}Setting Environment Variables
macOS/Linux (.bashrc, .zshrc):
export GITHUB_TOKEN="ghp_your_token_here"export DATABASE_URL="postgresql://user:pass@localhost:5432/db"export API_KEY="your-api-key"
Windows (PowerShell):
$env:GITHUB_TOKEN = "ghp_your_token_here"$env:DATABASE_URL = "postgresql://user:pass@localhost:5432/db"
Session-specific:
# Set variable before running Claude CodeGITHUB_TOKEN=token DATABASE_URL=url claudeAuthentication
OAuth 2.0 (Recommended)
Many cloud-based MCP servers use OAuth 2.0:
Process:
- Add server (no credentials needed initially)
- Start Claude Code
- Type
/mcpcommand - Follow browser login flow
- Tokens stored securely and auto-refreshed
Example:
# Add GitHub serverclaude mcp add --transport http github https://mcp.github.com# In Claude Code session/mcp# Opens browser for GitHub authentication
API Token Authentication
For services using API tokens:
Via header:
claude mcp add --transport http service https://api.example.com/mcp \ --header "Authorization: Bearer your-token"Via environment variable:
# Set environment variableexport SERVICE_TOKEN="your-token"# Add server with variable referenceclaude mcp add --transport http service https://api.example.com/mcp \ --header "Authorization: Bearer ${SERVICE_TOKEN}"
No Authentication
Public or local servers:
claude mcp add --transport http public https://public-api.example.com/mcp
Server Management Commands
List All Servers
claude mcp list
Output:
Configured MCP servers: github (http) - https://mcp.github.com filesystem (stdio) - npx -y @modelcontextprotocol/server-filesystem postgres (stdio) - npx -y @modelcontextprotocol/server-postgresGet Server Details
claude mcp get github
Output:
{ "type": "http", "url": "https://mcp.github.com", "headers": { "Authorization": "Bearer ..." }}Remove Server
claude mcp remove github
Reset Project Approval Choices
claude mcp reset-project-choices
Clears your approval decisions for project-scoped servers.
Check Server Status (In Claude Code)
/mcpShows:
- Connected servers
- Authentication status
- Available tools and resources
Using MCP Resources
Reference Resources with @
Type @ in Claude Code to see available resources:
Format:
@server:protocol://resource/pathExamples:
GitHub issue:
Review the implementation in @github:issue://123Notion page:
Summarize @notion:page://abc123def456File from filesystem server:
Compare @filesystem:file:///project/src/old.js with current implementationDatabase table:
Analyze schema for @postgres:table://usersUsing MCP Prompts
Execute Prompts with /
Type / to discover MCP prompts:
Format:
/mcp__servername__promptname [arguments]Examples:
GitHub PR review:
/mcp__github__review-pr 123Notion page creation:
/mcp__notion__create-page "Project Ideas" "Brainstorming session notes"Database query:
/mcp__postgres__query "SELECT * FROM users WHERE active = true"Output Limits
Default Limits
- Warning threshold: 10,000 tokens
- Default maximum: 25,000 tokens
Configure Output Limit
# Allow larger outputs (50,000 tokens)export MAX_MCP_OUTPUT_TOKENS=50000claude# Or inlineMAX_MCP_OUTPUT_TOKENS=50000 claude
Use when:
- Server returns large datasets
- Processing extensive documents
- Analyzing comprehensive logs
Timeout Configuration
MCP Server Startup Timeout
# 10-second timeout (default is 5 seconds)MCP_TIMEOUT=10000 claudeUse when:
- Server initialization is slow
- Network latency is high
- Complex server setup required
Popular MCP Servers
Official MCP Servers
GitHub:
claude mcp add --transport http github https://mcp.github.com
Tools: Create issues, review PRs, manage repositories
Filesystem:
claude mcp add --transport stdio filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/DocumentsTools: Read/write files, search directories
PostgreSQL:
claude mcp add --transport stdio postgres \ --env POSTGRES_CONNECTION_STRING=postgresql://localhost:5432/db \ -- npx -y @modelcontextprotocol/server-postgres
Tools: Query database, analyze schema
Slack:
claude mcp add --transport http slack https://mcp.slack.com
Tools: Send messages, search history, manage channels
Google Drive:
claude mcp add --transport http gdrive https://mcp.google.com/drive
Tools: Access files, search documents, manage folders
Brave Search:
claude mcp add --transport stdio brave-search \ --env BRAVE_API_KEY=your-key \ -- npx -y @modelcontextprotocol/server-brave-search
Tools: Web search, fact-checking
Git:
claude mcp add --transport stdio git -- npx -y @modelcontextprotocol/server-git
Tools: Repository operations, commit history, diff analysis
Sentry:
claude mcp add --transport http sentry https://mcp.sentry.io \ --header "Authorization: Bearer ${SENTRY_TOKEN}"Tools: Error monitoring, stack trace analysis
Linear:
claude mcp add --transport http linear https://mcp.linear.app
Tools: Issue tracking, project management
Notion:
claude mcp add --transport http notion https://mcp.notion.com/mcp
Tools: Page creation, database queries, content search
Community MCP Servers
Airtable:
claude mcp add --transport stdio airtable \ --env AIRTABLE_API_KEY=your-key \ -- npx -y airtable-mcp-server
MongoDB:
claude mcp add --transport stdio mongodb \ --env MONGODB_URI=mongodb://localhost:27017 \ -- npx -y mongodb-mcp-server
Redis:
claude mcp add --transport stdio redis \ --env REDIS_URL=redis://localhost:6379 \ -- npx -y redis-mcp-server
AWS:
claude mcp add --transport stdio aws \ --env AWS_PROFILE=default \ -- npx -y aws-mcp-server
Enterprise MCP Configuration
Managed MCP Servers
Administrators can deploy centralized configurations:
macOS:
/Library/Application Support/ClaudeCode/managed-mcp.jsonWindows:
C:\ProgramData\ClaudeCode\managed-mcp.jsonLinux:
/etc/claude-code/managed-mcp.jsonAllowlists/Denylists
managed-settings.json:
{ "mcpServers": { "allowlist": ["github", "slack", "notion"], "denylist": ["unapproved-server"] }}Controls which servers users can configure.
Security Considerations
Trust Verification
CRITICAL WARNING: "Use third party MCP servers at your own risk - Anthropic has not verified the correctness or security of all these servers."
Before installing:
- Verify server source (official vs community)
- Review server code if open source
- Check reputation and reviews
- Understand data access requirements
- Review authentication requirements
Prompt Injection Risks
Servers that fetch untrusted content (web search, user input) can expose Claude to prompt injection attacks.
Mitigation:
- Use servers from trusted sources
- Limit server permissions
- Review server behavior regularly
- Monitor for unexpected actions
Credential Management
Best practices:
- Use environment variables for secrets
- Never commit tokens to git (.mcp.json with ${VAR})
- Rotate credentials regularly
- Use least-privilege access tokens
- Enable OAuth when available
Bad practice:
{ "mcpServers": { "api": { "headers": { "Authorization": "Bearer hardcoded-token-123" } } }}Good practice:
{ "mcpServers": { "api": { "headers": { "Authorization": "Bearer ${API_TOKEN}" } } }}Project-Scoped Server Approval
Claude Code prompts before using servers from .mcp.json:
Prompt:
This project wants to connect to the following MCP servers:- github (https://mcp.github.com)- filesystem (local file access)Do you trust these servers?[Allow for this session] [Always allow] [Never allow]Review before approving:
- Server purpose and source
- Data access requirements
- Team trustworthiness
- Security implications
Troubleshooting
Server Not Appearing
Check configuration:
claude mcp listclaude mcp get server-name
Verify:
- Server name is correct
- Configuration syntax is valid
- Environment variables are set
- Server is running (stdio)
Authentication Failing
For OAuth:
/mcpRe-authenticate through browser
For API tokens:
- Verify environment variable is set:
echo $TOKEN_NAME - Check token hasn't expired
- Verify token permissions
Connection Timeout
Increase timeout:
MCP_TIMEOUT=15000 claude
Check:
- Network connectivity
- Server URL is correct
- Server is running and accessible
Tools Not Available
Verify server is connected:
/mcpCheck:
- Server status shows as connected
- Authentication is complete
- Server supports expected tools
Large Output Truncated
Increase output limit:
MAX_MCP_OUTPUT_TOKENS=50000 claude
Alternative:
- Request smaller data chunks
- Use server filtering options
- Paginate results
Windows npx Issues
Use cmd /c wrapper:
claude mcp add --transport stdio server -- cmd /c npx -y package-name
Testing MCP Servers
1. Verify Server Added
claude mcp list
Confirm server appears in the list.
2. Check Server Details
claude mcp get server-name
Verify configuration is correct.
3. Test in Claude Code
Start Claude Code and run:
/mcpConfirm:
- Server shows as connected
- Authentication status is OK
- Tools are available
4. Test Resources
Type @ and verify server resources appear:
@server-name:5. Test Prompts
Type / and verify server prompts appear:
/mcp__server-name__6. Execute Test Command
Try a simple server operation:
/mcp__github__list-reposCommon Use Cases
Development Workflow
GitHub + Filesystem + Git:
# GitHub for issues and PRsclaude mcp add --transport http github https://mcp.github.com# Filesystem for project accessclaude mcp add --transport stdio filesystem -- npx -y @modelcontextprotocol/server-filesystem ~/projects# Git for version controlclaude mcp add --transport stdio git -- npx -y @modelcontextprotocol/server-git
Usage:
- Review code from @filesystem:file://
- Reference issues with @github:issue://
- Check git history with /mcp__git__log
Database Administration
PostgreSQL + MongoDB:
# PostgreSQLclaude mcp add --transport stdio postgres \ --env POSTGRES_CONNECTION_STRING=${DATABASE_URL} \ -- npx -y @modelcontextprotocol/server-postgres# MongoDBclaude mcp add --transport stdio mongodb \ --env MONGODB_URI=${MONGO_URL} \ -- npx -y mongodb-mcp-server
Usage:
- Query data:
/mcp__postgres__query "SELECT ..." - Analyze schema: @postgres:schema://table_name
- Monitor collections: @mongodb:collection://users
Project Management
Linear + Slack + Notion:
# Linear for issuesclaude mcp add --transport http linear https://mcp.linear.app# Slack for communicationclaude mcp add --transport http slack https://mcp.slack.com# Notion for documentationclaude mcp add --transport http notion https://mcp.notion.com/mcp
Usage:
- Create issues:
/mcp__linear__create-issue - Send updates:
/mcp__slack__send-message - Reference docs: @notion:page://abc123
Error Monitoring
Sentry + Logs:
# Sentry for error trackingclaude mcp add --transport http sentry https://mcp.sentry.io \ --header "Authorization: Bearer ${SENTRY_TOKEN}"# Custom log serverclaude mcp add --transport stdio logs -- python3 ~/mcp-servers/log-analyzer.py
Usage:
- Analyze errors: @sentry:issue://ERROR-123
- Search logs:
/mcp__logs__search "authentication failed"
Best Practices Checklist
When connecting an MCP server:
- Server source is trusted and verified
- Transport type matches use case (HTTP for remote, stdio for local)
- Scope is appropriate (local/project/user)
- Credentials use environment variables (not hardcoded)
- Authentication is configured (OAuth or API token)
- Server name is descriptive and unique
- Environment variables are set before adding server
- Configuration is tested (
claude mcp list,/mcp) - Resources and prompts are accessible (@ and /)
- Team is informed (for project-scoped servers)
- Security implications reviewed
- .mcp.json is committed to git (project scope)
- Secrets are NOT committed to git
Key Principles
- Trust first - Only install servers from verified sources
- Environment variables - Use ${VAR} for secrets, never hardcode
- Scope appropriately - Local for personal, project for team, user for cross-project
- OAuth when possible - Prefer OAuth over API tokens for security
- Test thoroughly - Verify connection and tools before relying on server
- Document for team - Explain server purpose in project README
- Monitor access - Review what data servers can access
- Keep updated - Update servers regularly for security patches
- Least privilege - Grant minimum necessary permissions
- Review regularly - Audit configured servers periodically
Workflow Summary
When user asks to connect an MCP server:
- Identify server - Which service/integration is needed?
- Find installation method - Official docs, GitHub, or custom
- Choose transport - HTTP (remote) or stdio (local)?
- Gather credentials - API keys, OAuth setup, connection strings
- Set environment variables - Export secrets before adding server
- Select scope - Local (default), project (team), or user (cross-project)
- Add server - Use
claude mcp addcommand - Verify configuration -
claude mcp listandclaude mcp get - Test connection -
/mcpin Claude Code session - Use server - Access resources (@) and prompts (/)
Remember: MCP servers extend Claude's capabilities by connecting to external services. Always verify server trust, use environment variables for secrets, and test thoroughly before relying on server integrations.





首頁
