Creativity · MCP — client
Claude Desktop as an MCP Client
Claude Desktop is the canonical Model Context Protocol client. It runs Claude the assistant on your machine and loads MCP servers defined in a claude_desktop_config.json file, spawning each one as a child process over stdio. It is the easiest place to start with MCP — add a server block, restart Claude, and the tools appear in chat.
MCP facts
- Kind
- client
- Ecosystem
- anthropic-mcp
- Language
- native (Electron)
- Transports
- stdio, http, sse
Capabilities
- Loads MCP servers from claude_desktop_config.json on startup
- Supports stdio (child process) and remote HTTP / SSE transports
- Tool-call UI: Claude asks permission before calling a tool the first time
- Ships on macOS + Windows; Linux via community builds only
Install
Download from https://claude.ai/download Configuration
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents"]
},
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx" }
}
}
} Frequently asked questions
Where do I put claude_desktop_config.json?
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json. Windows: %APPDATA%\Claude\claude_desktop_config.json. Edit, save, then restart Claude Desktop — it reloads the file on launch.
How do I debug an MCP server that does not show up?
Open Developer → Show Logs (or check ~/Library/Logs/Claude/mcp*.log on macOS). Missing Node/npx, bad env vars, or syntax errors in the JSON are the usual causes. Also try running the exact command/args from a terminal.
Is the tool-call UI safe?
Claude Desktop prompts for consent on first use of each tool and re-prompts on suspicious invocations. You can also pre-approve tools per server. Treat tool consent as a security boundary and read the invocation before clicking Allow.
Sources
- Claude Desktop — MCP quickstart — accessed 2026-04-20
- Claude Desktop download — accessed 2026-04-20