There are thousands of MCP servers on GitHub right now. Most of them are abandoned experiments with 12 stars. The awesome-mcp-servers list alone has 85,000 stars, which tells you how many people are looking, and how little curation they find once they arrive.
MCPFast tracks MCP releases across GitHub, Reddit, Hacker News and X, every two hours. This list comes from that data: the servers below are ranked by GitHub stars as of August 2026, but stars only got them on the list. What kept them on is that people actually run them.
One note on method. Stars measure attention, not quality. A server can trend for a week because of one viral post and be unusable in practice. So for each entry I looked at what the server does in a real workflow, and I say who should skip it. If you want the live version of this ranking, it updates hourly on our tools page.
1. modelcontextprotocol/servers (83,000 stars)
The reference implementations maintained alongside the protocol itself. Filesystem, fetch, git, memory, time. None of them are exciting, all of them are solid.
This is where to start if you have never touched MCP. Read the filesystem server's source before writing your own: it is short, and it shows the patterns the SDK expects. Skip it if you already have a working setup, since these servers are deliberately minimal.
2. czlonkowski/n8n-mcp (21,800 stars)
n8n-mcp lets an LLM build n8n workflows from a text description. You describe the automation, it assembles the nodes.
The interesting part is that it inverts the usual pitch. Instead of an AI agent doing the work, the AI builds the machine that does the work, and the machine is inspectable n8n, not a black box. If you already run n8n, this is worth an afternoon. If you don't, it won't convince you to start.
3. GLips/Figma-Context-MCP (15,100 stars)
Figma-Context-MCP feeds Figma layout data to coding agents like Cursor or Claude Code. The agent stops guessing what your design looks like and reads the actual frame structure.
Front-end developers get the most out of this one. The gap between "build this from a screenshot" and "build this from the Figma file" is bigger than you'd expect, mostly in spacing and component boundaries.
4. zilliztech/claude-context (10,100 stars)
claude-context indexes your whole codebase so Claude can search it semantically instead of grepping file by file. Built by the team behind the Milvus vector database.
Useful past a certain repo size, maybe 100k lines. Below that, plain grep and a well-written CLAUDE.md do the same job with zero moving parts. Above that, semantic search starts paying for its setup cost.
5. mcp-use/mcp-use (9,800 stars)
mcp-use is a framework rather than a server: it helps you build MCP servers and connect agents to them, in Python and TypeScript.
I hesitated to include it since it's not a server you install. But if this list makes you want to write your own server, this is the fastest route I know, and that seems worth a slot.
6. awslabs/mcp (9,000 stars)
AWS's official MCP servers. Cost analysis, CDK guidance, documentation lookup, Bedrock, a couple dozen more.
The quality varies between the individual servers, which is normal for a repo that is really twenty repos. The cost-analysis one alone justifies the install if your AWS bill is a recurring surprise.
7. firecrawl/firecrawl-mcp-server (6,500 stars)
Firecrawl's server gives your agent web scraping and search. Pages come back as clean markdown instead of raw HTML soup.
This is the one I'd pick for research-heavy workflows. Be aware it's a hosted service with a free tier, not a fully local tool, so your scraped URLs transit through their API.
8. wonderwhy-er/DesktopCommanderMCP (6,000 stars)
DesktopCommander gives Claude Desktop terminal control, file search and diff-based editing. It essentially turns the chat app into a coding agent.
If you live in Claude Code or Cursor, you already have all of this. It matters for people who work from Claude Desktop and want it to touch their machine, which is a real audience, just not every audience.
9. getsentry/XcodeBuildMCP (5,800 stars)
XcodeBuildMCP, maintained by Sentry, wraps Xcode builds, simulators and device management for agents. iOS development has always been the place where coding agents hit a wall of GUI-only tooling. This chips away at that wall.
Niche by definition. If you ship iOS or macOS apps, it might be the most useful entry on this page.
10. mksglu/context-mode (6,400 stars)
context-mode sits between your agent and its MCP tools and virtualizes the context they consume. Tool definitions are getting fat, and agents waste tokens carrying schemas they never call. This is an early attempt at fixing that.
I'd call it promising rather than proven. The problem it targets is real and getting worse, so it earns the last spot over safer picks.
What didn't make the list
Platforms like n8n, Dify or LibreChat show up in every MCP conversation because they speak the protocol, but they are products that consume MCP, not servers you add to an agent. Same for the official TypeScript SDK: essential if you build servers, irrelevant if you only use them.
The honest summary: install nothing until a task annoys you twice. Then check whether one of these ten fixes it. Adding six servers on day one mostly buys you a slower agent and a longer config file.
New servers ship every week. The MCPFast feed catches them within two hours, and the weekly roundup covers the ones that got traction. If you work in Claude Code specifically, we wrote a separate guide for that setup.