MCP Example Integrations

Claude Desktop config (stdio transport):

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "<from-vault>"
      }
    }
  }
}

Tool call flow:

1. User: "Review PR #142 in anthropics/claude-code"
2. Claude calls get_pull_request(owner="anthropics", repo="claude-code", pull_number=142)
3. GitHub MCP server returns PR metadata (title, description, author, files changed)
4. Claude calls get_pull_request_files(owner, repo, pull_number) to get the diff
5. Claude generates review based on PR description + diff, no copy-paste required

Tool call flow:

1. Claude calls list_tables() to discover available tables
2. Claude calls describe_table(table="orders") to understand the schema
3. Claude generates a SQL query and calls query(sql="SELECT region, SUM(amount)...")
4. Claude interprets results and flags regions where sales deviated >20% from prior month

Production safeguards:

• Database user is readonly — no INSERT, UPDATE, DELETE permissions
• Row-level security filters data to the current user's accessible records
• Query timeout set at 30 seconds — prevents accidental full-table scans
• Result size capped at 10,000 rows — prevents context window overflow from large result sets
• Sensitive columns (PII) are masked or excluded from the database user's view

Claude Desktop config (scoped to specific directory):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/home/user/research-papers"
      ]
    }
  }
}

The path argument scopes the server to that directory only — the AI cannot read files outside it.

Tool call flow:

1. User: "What papers discuss transformer architectures from 2023?"
2. Claude calls list_directory(path="/") to see available files
3. Claude calls search_files(pattern="transformer") or reads relevant filenames
4. Claude calls read_file(path="attention-is-all-you-need.pdf") for promising papers
5. Claude synthesises across read papers to answer the question

Tool call flow:

1. Claude calls list_channels() to confirm the channel name
2. Claude calls read_channel(channel="#engineering", since="7d")
3. Slack MCP server returns message history (user display names, not IDs)
4. Claude synthesises key decisions, blockers, and completed work
5. Claude drafts a status report — user reviews before any send action