MCP Reference Servers 🔌 MCP server Free

What it does

The modelcontextprotocol/servers repository contains the reference servers maintained by the MCP steering group. They exist mainly to show how the protocol and the official SDKs work, and several are also handy in everyday use. The repository used to hold many more servers, for services such as GitHub, Slack, PostgreSQL, Puppeteer and Brave Search. Those have been moved to a separate archive repo, and several were replaced by official servers from the vendors themselves. To browse MCP servers in general, the maintainers now point to the official MCP Registry. The project is moving from the MIT licence to Apache-2.0.

Tools it gives your agent

The seven current reference servers:

• Everything: a test server that exercises prompts, resources and tools
• Fetch: fetches web content and converts it for LLM use
• Filesystem: reads, writes and searches files within directories you allow
• Git: reads, searches and changes a local Git repository
• Memory: persistent memory based on a knowledge graph
• Sequential Thinking: structured, step-by-step problem solving
• Time: time and time zone conversion

Works with

They run locally over stdio, so any MCP client that can start a command works, including Claude Code, Claude Desktop, Cursor, Codex, Gemini CLI and VS Code. The TypeScript servers run through npx. The Python ones (Fetch, Git and Time) run through uvx or pip.

How to connect

Examples from the README:

npx -y @modelcontextprotocol/server-memory
uvx mcp-server-git

In a client config, with the Filesystem server limited to one folder:

{ "mcpServers": { "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/files"] } } }

Cost and limits

They are all free and open source, and none needs an account or API key.

Safety notes

The maintainers describe these as educational reference implementations, not production-ready servers, and say you should do your own security review. Fetch can reach local and internal IP addresses, which the README flags as a risk. By default it respects robots.txt for requests the model makes. Filesystem only works inside the directories you pass or the roots your client provides, so keep that list as small as possible. Claude Code and similar agents already have file and git tools built in.