Model Context Protocol (MCP)

An open protocol that lets AI clients (Claude Code, Claude Desktop, IDEs, Open WebUI and others) call tools and read data exposed by servers. Use it to give an agent controlled access to files, Git hosts, databases and your homelab.

Documentation: modelcontextprotocol.io · Claude Code MCP docs · Reference servers

Verification: claude mcp commands were checked against claude --help. Protocol concepts and the config shape are from the official docs. Package names of third-party servers change; confirm each before use (marked (unverified)).

Concepts #

TERM	MEANING
Client / host	The app the user talks to (Claude Code, Claude Desktop)
Server	A process that exposes capabilities
Tools	Actions the model can call
Resources	Data the client can read
Prompts	Reusable prompt templates the server offers
Transport `stdio`	Client launches the server locally and talks over stdin and stdout
Transport streamable HTTP	Remote server over HTTP; the current remote transport
Transport SSE	Older remote transport, deprecated in favour of HTTP

Add Servers in Claude Code #

COMMAND	DESCRIPTION
`claude mcp add --transport http NAME URL`	Remote HTTP server
`claude mcp add NAME -e KEY=VALUE -- COMMAND ARGS`	Local stdio server (everything after `--` is the command)
`claude mcp add -s project NAME ...`	Write to `.mcp.json` so the team shares it
`claude mcp add-json NAME 'JSON'`	From a JSON definition
`claude mcp list` / `claude mcp get NAME`	Inspect and health check
`claude mcp login NAME`	Complete OAuth for a server
`claude mcp remove NAME`	Remove
`/mcp` (in a session)	Status, authenticate, enable or disable

Scopes: local (default, you in this project), project (.mcp.json, shared), user (all your projects).

Config File #

.mcp.json in the project root (Claude Code) and the mcpServers block in Claude Desktop’s config use the same shape.

json
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/srv/DATA_DIR"]
    },
    "example-remote": {
      "type": "http",
      "url": "https://mcp.example.com/mcp"
    },
    "example-local": {
      "command": "/usr/local/bin/SERVER_BINARY",
      "env": { "API_TOKEN": "${API_TOKEN}" }
    }
  }
}

Reference environment variables rather than pasting tokens. Do not commit secrets in .mcp.json.

Common Servers #

SERVER	PURPOSE
`@modelcontextprotocol/server-filesystem`	Read and write a limited set of directories
`@modelcontextprotocol/server-memory`	Simple knowledge-graph memory `(unverified)`
`@modelcontextprotocol/server-fetch`	Fetch web pages `(unverified: the reference fetch server is Python, run through uvx)`
GitHub’s MCP server	Issues, pull requests and repos `(unverified: see the github/github-mcp-server repo)`
Context7	Current library documentation `(unverified)`
Playwright MCP	Browser automation `(unverified)`

Safety #

A server runs with your privileges (stdio) or holds your tokens (remote). Install only servers you trust and pin versions.
Restrict filesystem servers to specific directories, not / or $HOME.
Prefer read-only credentials. Tool output can contain prompt injection, so keep write access narrow.
Debug a server with the MCP Inspector: npx @modelcontextprotocol/inspector (unverified).

See Also #

Updated 7 Nov 2026 Edit this page on GitHub