Model Context Protocol (MCP) lets Claude access local tools, files, and APIs. Instead of pasting context into chat, you give Claude direct access to what it needs.
Setup with Claude Code
Add servers to ~/.claude/mcp.json (global) or .mcp.json (per-project):
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/dir"]
}
}
}Svelte Development
The Svelte MCP server gives Claude access to current Svelte 5 docs, migration guides, and API references – so it writes modern Svelte instead of outdated Svelte 4 patterns.
{
"mcpServers": {
"svelte": {
"command": "npx",
"args": ["-y", "svelte-llm", "mcp"]
}
}
}Useful MCP Servers
- filesystem - Read/write files in allowed directories
- git - Git operations (log, diff, blame)
- sqlite - Query local databases
- brave-search - Web search from Claude
- postgres - Query Postgres databases
Full directory: github.com/modelcontextprotocol/servers
Why this matters
Without MCP, you copy-paste code into Claude and hope it has enough context. With MCP, Claude reads your actual files, queries your actual database, and checks the actual docs. The answers are better because the context is real.