The top 10 MCP servers in 2026, ranked by GitHub stars on 17 November 2026, are Context7 (65,011), Chrome DevTools MCP (58,007), codebase-memory-mcp (48,774), Playwright MCP (39,799), MCP for Blender (34,801), GitHub MCP Server (34,546), Serena (31,481), n8n-MCP (23,275), xiaohongshu-mcp (17,182) and MCP Toolbox for Databases (16,897). The most popular MCP server by weekly npm downloads is Playwright MCP, installed 10,698,162 times in a week, four times as often as Chrome DevTools MCP and more than nine times as often as Context7.
This ranking uses GitHub stars — a measure of attention, not of use. Every figure below links to the API record it was read from, so you can check any of them. Where a server publishes an npm package, its weekly download count is shown beside the star count, because the two disagree in ways that are the most useful thing on this page.
Which MCP servers have the most GitHub stars in 2026?
| Rank | Server | GitHub stars | npm downloads/week | Language | What it does |
|---|---|---|---|---|---|
| 1 | Context7 | 65,011 | 1,166,420 | TypeScript | Current library docs and code examples |
| 2 | Chrome DevTools MCP | 58,007 | 2,648,086 | TypeScript | Chrome DevTools: debugging, network, performance |
| 3 | codebase-memory-mcp | 48,774 | 9,610 | C | Persistent code knowledge graph |
| 4 | Playwright MCP | 39,799 | 10,698,162 | TypeScript | Browser control via accessibility tree |
| 5 | MCP for Blender | 34,801 | not published | Python | 3D modelling and scenes in Blender |
| 6 | GitHub MCP Server | 34,546 | not published | Go | Repos, issues, pull requests, Actions |
| 7 | Serena | 31,481 | not published | Python | Semantic code retrieval and editing |
| 8 | n8n-MCP | 23,275 | 41,928 | TypeScript | Build and debug n8n workflows |
| 9 | xiaohongshu-mcp | 17,182 | not published | Go | Search and post on Xiaohongshu (RedNote) |
| 10 | MCP Toolbox for Databases | 16,897 | 54,434 | Go | Database queries and tools |
Star counts were read on 17 November 2026 at 20:43 UTC. npm counts cover the week 2026-11-10 to 2026-11-16. Both move daily, so treat the table as a dated snapshot, not a standing fact.
Where can you verify each server's numbers?
Each link returns the raw record the figure came from.
- Context7 — repo · npm
- Chrome DevTools MCP — repo · npm
- codebase-memory-mcp — repo · npm
- Playwright MCP — repo · npm
- MCP for Blender — repo
- GitHub MCP Server — repo
- Serena — repo
- n8n-MCP — repo · npm
- xiaohongshu-mcp — repo
- MCP Toolbox for Databases — repo · npm
How were these 10 MCP servers selected?
A ranking is only as honest as its candidate list, so the list was not picked by hand. It was built by a rule you can rerun:
- Find every candidate. Search GitHub, sorted by stars, for repositories tagged with the
mcp-servertopic, and for repositories with "MCP server" in their name or description. That gave 514 active repositories. The topic alone is not enough: Playwright MCP and MCP for Blender do not carry it. - Keep a repository if it names itself an MCP server. It qualifies if it has the
mcp-servertopic, has "mcp" in its repository name, or calls itself an MCP server in its description. - Drop what is not a server you add to an MCP client. Three kinds of repository pass step 2 but are not servers, and they are listed below.
Directories are not servers. The two most-starred repositories in this space are catalogues: punkpeye/awesome-mcp-servers at 97,818 stars and modelcontextprotocol/servers at 92,075. Both outrank every entry in the table. Listing them first would be like ranking a bibliography above the papers in it.
SDKs and frameworks are not servers either. PrefectHQ/fastmcp has 28,633 stars, modelcontextprotocol/python-sdk 24,885 and modelcontextprotocol/typescript-sdk 13,778, with 90,522,817 weekly downloads on @modelcontextprotocol/sdk. That download figure dwarfs every server here, and it should: it is the library most of them are built on.
Apps that also ship an MCP server are left out. Large applications such as n8n and Gemini CLI carry the mcp-server topic because they can speak MCP, but the thing you install is the application. next-ai-draw-io, at 37,024 stars, describes itself as "Web app, desktop app, and an MCP server", so it is out for the same reason.
Which repositories were close calls?
Two repositories could reasonably go either way, and each would enter the top 10 if counted:
- context-mode, 34,230 stars, carries the
mcp-servertopic but describes itself as "a free, local plugin for 11 coding agents". Counted, it would be seventh. - XERJ, 19,200 stars, calls itself an "AI search engine and MCP server": the engine comes first. Counted, it would be ninth.
Both were left out because the thing you install is a plugin or a search engine, with MCP as one way in. If you count both, Serena drops to eighth and n8n-MCP to ninth, and xiaohongshu-mcp and MCP Toolbox for Databases fall out of the top 10.
Why do stars and downloads rank these servers so differently?
Because they measure different things. A star is one person signalling interest once. A download is a machine installing the package, repeatedly, often inside CI or every time an agent starts a server with npx.
The six top-10 servers that publish to npm rank like this:
| Server | Rank by stars | Rank by downloads | npm downloads/week |
|---|---|---|---|
| Playwright MCP | 4 | 1 | 10,698,162 |
| Chrome DevTools MCP | 2 | 2 | 2,648,086 |
| Context7 | 1 | 3 | 1,166,420 |
| MCP Toolbox for Databases | 10 | 4 | 54,434 |
| n8n-MCP | 8 | 5 | 41,928 |
| codebase-memory-mcp | 3 | 6 | 9,610 |
Playwright MCP is the clearest case. It is fourth by stars, and its 10,698,162 weekly downloads are more than the other five combined. A browser server that agents start on demand is installed far more often than it is starred.
codebase-memory-mcp runs the other way. Third by stars at 48,774, it has the fewest downloads of the six at 9,610. It is written in C and its main install path is a native binary from a shell script, so the npm package is one of several ways to get it. Its repository was created on 2026-02-24, the newest in the top 10.
Figma Context MCP, outside the top 10, shows the same gap. It is twelfth by stars at 16,232, but its 144,185 weekly downloads would place it fourth among the eight servers in this list that publish to npm.
Four of the top 10 publish nothing to npm: GitHub MCP Server and xiaohongshu-mcp ship Go binaries and containers, and MCP for Blender and Serena are Python packages. An empty download column means npm is the wrong place to count them, not that nobody uses them.
What does each of the top 10 actually do?
1. Context7 — 65,011 stars
Serves current documentation and code examples for a library into the model's context, so an agent writing against a framework gets that framework's present API rather than a recalled one. Created 2025-03-26. 1,166,420 weekly downloads.
2. Chrome DevTools MCP — 58,007 stars
The Chrome DevTools team's server. It lets a coding agent control and inspect a live Chrome browser for automation, debugging and performance analysis. It is also available as a CLI. Created 2025-09-11, the second-newest in the top 10, and 2,648,086 weekly downloads.
3. codebase-memory-mcp — 48,774 stars
Builds a persistent local knowledge graph of a repository, from source code to documentation, configuration and infrastructure, and exposes it to agents as MCP tools. Its README calls the MCP server "the core of the project". Written in C.
4. Playwright MCP — 39,799 stars
Microsoft's browser automation server. It drives a browser through its accessibility tree rather than through screenshots, so an agent works on page structure instead of pixels. 10,698,162 weekly downloads, the most of any server here.
5. MCP for Blender — 34,801 stars
Prompt-driven 3D modelling, scene creation and manipulation in Blender. It was formerly named blender-mcp; it is a third-party integration, not made by the Blender Foundation. Distributed on PyPI.
6. GitHub MCP Server — 34,546 stars
GitHub's own server for repositories, issues, pull requests and Actions. Written in Go and distributed as a binary and a container rather than on npm. 5,362 forks, the most of any server here.
7. Serena — 31,481 stars
Semantic code retrieval, editing and refactoring tools that work at the level of symbols rather than lines, "the IDE for your coding agent". Written in Python.
8. n8n-MCP — 23,275 stars
Lets an AI assistant operate n8n: read failed executions, edit workflows in place, validate them and build new ones. 41,928 weekly downloads.
9. xiaohongshu-mcp — 17,182 stars
Gives an AI assistant access to Xiaohongshu (RedNote): search, read and publish posts. Written in Go, and the only server in the top 10 built around a single consumer platform.
10. MCP Toolbox for Databases — 16,897 stars
Google's open-source MCP server for databases. It works both as a ready-made server for agents and IDEs and as a framework for defining safe custom database tools. Created 2024-06-07, the oldest repository in the top 10. 54,434 weekly downloads.
Which servers just missed the top ten?
- Apify MCP Server, 16,706 stars, 191 behind MCP Toolbox for Databases. Web scraping through Apify Actors; 48,662 weekly downloads.
- Figma Context MCP, 16,232 stars. Figma layout data for coding agents; 144,185 weekly downloads, more than MCP Toolbox for Databases and n8n-MCP combined.
- Unity MCP, 16,037 stars. Assets, scenes and scripts in the Unity Editor, written in C#.
Positions nine to thirteen are separated by little more than a thousand stars, so this end of the list can reorder within weeks.
What do stars, forks and weekly downloads actually measure?
| Measure | What it counts, and what it misses |
|---|---|
| GitHub stars | One signal of interest per account, kept forever. Counts attention at any point since creation; never decays when a project is abandoned. |
| Forks | Copies made to modify. High forks against stars usually means a tool people adapt rather than adopt. |
| npm downloads/week | Installs by machines, often CI or npx starting a server, counted over seven days. Reflects current use, but misses anything distributed as a binary, a container or a Python package. |
| Last pushed | The most recent commit. The cheapest check on whether a star count is describing something still alive. |
What are the limits of ranking MCP servers this way?
- Stars never go down. A server that was popular in 2025 and abandoned in 2026 keeps its count. All ten servers here were pushed to within the week the figures were read.
- npm is one channel. Go binaries, Docker images and PyPI packages are invisible to it. Four of the top 10 show nothing in the download column.
- Selection is a rule, and rules have edges. The two close calls above would change positions seven to ten. The rule is stated so you can disagree with it specifically.
- Both counts are snapshots. They were read on 17 November 2026 and were already changing while being read.
- Neither measures quality. Nothing here says which server is well built, well documented or safe to give write access to.
How is this different from being cited by an AI search engine?
Stars and downloads tell you how many people adopted a tool. They tell you nothing about whether an AI assistant will name it when someone asks "what's the best MCP server for browser automation". That is a different kind of visibility, and it is measured differently: across a set of real queries, which products the assistants mention, and your share of voice against the competitors mentioned alongside you. If that is the question you care about, start with what AI visibility is or how to track brand mentions in AI search.
Frequently asked questions
By GitHub stars on 17 November 2026: Context7 (65,011), Chrome DevTools MCP (58,007), codebase-memory-mcp (48,774), Playwright MCP (39,799), MCP for Blender (34,801), GitHub MCP Server (34,546), Serena (31,481), n8n-MCP (23,275), xiaohongshu-mcp (17,182) and MCP Toolbox for Databases (16,897).
It depends on the measure. By GitHub stars, Context7 leads with 65,011. By weekly npm downloads for 2026-11-10 to 2026-11-16, Microsoft's Playwright MCP leads with 10,698,162, against 2,648,086 for Chrome DevTools MCP and 1,166,420 for Context7.
Because it is a directory of servers, not a server. It has 92,075 stars and punkpeye/awesome-mcp-servers has 97,818, so both would top the table, but ranking a catalogue above the things it catalogues would make the list mean something other than its title. SDKs and frameworks such as FastMCP are excluded for the same reason.
Among servers, Playwright MCP, at 10,698,162 weekly npm downloads. The @modelcontextprotocol/sdk package is far larger at 90,522,817, but it is the SDK most servers are built on rather than a server itself.
No. codebase-memory-mcp is third by stars at 48,774 but has 9,610 weekly npm downloads, while Playwright MCP is fourth by stars with 10,698,162. Stars count attention accumulated since a repository was created; downloads count recent installs through one channel.
It does not measure it. Adoption counts say how many developers installed a tool; AI search visibility is whether an assistant names your product when someone asks a buying question. The second kind is tracked across a set of queries, by which brands get mentioned and how your share of voice compares with the competitors mentioned beside you.
The short version
On 17 November 2026 the top 10 MCP servers by GitHub stars were Context7, Chrome DevTools MCP, codebase-memory-mcp, Playwright MCP, MCP for Blender, GitHub MCP Server, Serena, n8n-MCP, xiaohongshu-mcp and MCP Toolbox for Databases. Ranked by weekly npm downloads instead, Playwright MCP leads by four times. The list comes from a stated rule over 514 GitHub repositories, and every figure on this page links to the API record it came from.