MCP Reference Servers
by Model Context Protocol
Specification by example: seven small, correct MCP servers you can run with npx or uvx in seconds.
Overview
modelcontextprotocol/servers is the official reference-implementation repository for the Model Context Protocol, and at 89.3k stars and 11.4k forks it is the most-starred artifact in the MCP ecosystem — largely because it is the canonical answer to the question every new implementer has: what does a correct MCP server actually look like? It ships seven maintained reference servers, deliberately small and readable rather than feature-complete: Everything, a reference and test server exercising prompts, resources and tools together so client authors can validate their host; Fetch, web content fetching and conversion for efficient LLM usage; Filesystem, secure file operations with configurable access controls; Git, tools to read, search and manipulate Git repositories; Memory, a knowledge-graph-based persistent memory system; Sequential Thinking, dynamic and reflective problem-solving through thought sequences; and Time, time and timezone conversion. Roughly thirteen earlier servers — GitHub, GitLab, PostgreSQL, SQLite, Slack and others — are archived rather than deleted, which matters in practice because plenty of tutorials still link to them; those integrations are now maintained by their respective vendors. Implementations span TypeScript and Python, and both are runnable without cloning the repo: npx -y @modelcontextprotocol/server-memory for the Node servers, and uvx mcp-server-git or pip install mcp_server_git for the Python ones. The README is organised as Reference Servers, Getting Started, Creating Your Own Server, Learn More, Contributing, Releasing, Security, License, Community and Support, and it now directs discovery of third-party servers to the official MCP Registry rather than maintaining an in-repo community list. Licensed MIT for existing code, with new contributions under Apache-2.0.
At a Glance
- Topic
- MCP
- Level
- All Levels
- Format
- Documentation
- Cost
- Free
- Duration
- ~1 hour to run the reference servers; a few hours to read one implementation end to end
- Provider
- Model Context Protocol
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Read a minimal, correct MCP server implementation instead of inferring one from the specification
- ✓Run any reference server instantly with npx or uvx, with no clone or build step
- ✓See how tools, resources and prompts coexist in one server via the Everything test server
- ✓Implement configurable access controls the way the Filesystem server scopes its reach
- ✓Model persistent agent memory as a knowledge graph rather than a flat conversation transcript
- ✓Recognise which legacy servers are archived so you stop following stale tutorials
- ✓Discover third-party servers through the official MCP Registry instead of in-repo lists
- ✓Compare the TypeScript and Python idioms for the same protocol on equivalent servers
Highlights
- •89.3k stars — effectively specification-by-example for the entire MCP ecosystem
- •Seven servers kept deliberately small and readable, so a full implementation fits in one sitting
- •Both TypeScript and Python references, which is rare and lets you compare idioms directly
- •Zero-install execution through npx and uvx makes it a working demo, not just source to skim
- •The Everything server exists specifically to exercise clients — the fastest way to validate a host you are building
- •Archived servers stay visible rather than being deleted, which is honest about what moved to vendor ownership
Who It’s For
Best For
- ✓Developers writing their first MCP server who want a correct example to work from
- ✓Client and host authors who need a server that exercises every protocol capability
- ✓Engineers evaluating MCP before committing a team to it
- ✓Anyone wiring filesystem, git or persistent memory access into an AI coding assistant
Prerequisites
- •Node.js with npx available, or Python with uv/uvx installed
- •An MCP-capable client such as Claude Desktop or Claude Code to connect the servers to
- •Basic familiarity with MCP's tools, resources and prompts concepts
FAQ
What is MCP Reference Servers?
The official reference-implementation repository for the Model Context Protocol and the most-starred artifact in the ecosystem at 89.3k stars. It ships seven deliberately small servers in TypeScript and Python — Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time — that answer the question 'what does a correct MCP server actually look like?' and run with no clone or build step.
Is MCP Reference Servers free?
MCP Reference Servers is free to access.
What level is MCP Reference Servers for?
MCP Reference Servers is aimed at a all levels audience. Recommended background: Node.js with npx available, or Python with uv/uvx installed, An MCP-capable client such as Claude Desktop or Claude Code to connect the servers to, Basic familiarity with MCP's tools, resources and prompts concepts.
How long does MCP Reference Servers take?
Expect roughly ~1 hour to run the reference servers; a few hours to read one implementation end to end. Most learners work through it at their own pace.
What will I learn from MCP Reference Servers?
You'll learn: Read a minimal, correct MCP server implementation instead of inferring one from the specification; Run any reference server instantly with npx or uvx, with no clone or build step; See how tools, resources and prompts coexist in one server via the Everything test server; Implement configurable access controls the way the Filesystem server scopes its reach; Model persistent agent memory as a knowledge graph rather than a flat conversation transcript; Recognise which legacy servers are archived so you stop following stale tutorials; Discover third-party servers through the official MCP Registry instead of in-repo lists; Compare the TypeScript and Python idioms for the same protocol on equivalent servers.
Topics
Sources
This page was written from 3 sources, 2 on domains other than github.com.