TL;DR: This best MCP servers list ranks 250 Model Context Protocol servers from the official MCP Registry on real adoption data. Top of the list on our composite quality score is Repowise at 96/100; the most-downloaded is Chrome DevTools MCP at 1,051,748 weekly package downloads. 168 of 250 (67%) are first-party servers built by the vendor whose product they connect to, and every one carries copy-paste install snippets for 4 AI clients.
Last updated: 11 September 2026. Data pulled: 11 September 2026, from the official MCP Registry plus the GitHub API, npm, PyPI and the GitHub Advisory Database. Servers ranked: 250. Combined weekly package downloads: 3,897,997.
Data note. GitHub enrichment covers 246 of the 250 servers in this build, so star counts, licences and release history are reported directly from the GitHub API.
An MCP server is a program that exposes tools, data and prompts to AI applications through the Model Context Protocol, the open standard Anthropic released in November 2024. It is the adapter that lets Claude, Cursor, VS Code or ChatGPT read from and act on an external system without integration code written for that one system. The same server works across AI apps, which is the whole reason LLMs converged on one protocol instead of a connector per vendor.
Most MCP lists are a copy of somebody’s GitHub “awesome” file. This one is built from the official registry and enriched with adoption, maintenance and security data, so you can compare servers rather than just discover them.
The numbers worth quoting
- 250 MCP servers ranked as of 11 September 2026 across 11 categories. 57 of them (23%) ship an MCP-specific npm or PyPI package, together drawing 3,897,997 downloads a week.
- Chrome DevTools MCP is the most downloaded MCP server at 1,051,748 a week, 27.0% of attributable download volume. The top three take 56.7% and the top ten 86.8%.
- 26 servers are excluded from those download totals because their package is a wider library rather than the MCP server, setting aside 4,015,103 weekly downloads (51% of the raw sum). Most MCP rankings sum these in.
- 168 of 250 (67%) are first-party, published under a registry-verified namespace the vendor owns. They account for 82.1% of download volume.
- Developer Tools is the biggest category at 81 servers (32%), followed by Search & Web at 50.
- 162 of 250 run over stdio as a local process. Only 109 (44%) offer Streamable HTTP, and just 93 publish a hosted remote endpoint.
- All 250 servers ship install snippets for Claude Desktop, Cursor, VS Code and Claude Code, so adding any of them is a copy-paste.
- 99 publish to npm and 58 to PyPI, which is why Node and Python are effectively the two runtimes of the MCP ecosystem.
- 52 security advisories on record across 17 of the 250 servers. 126 servers (50%) have no known advisory, 15 carry only patched ones, two carry an unpatched one and are named in the security section below, and 107 cannot be checked at all because they publish no queryable package.
- Maintenance is strong: median 1 days since the last update, 227 servers (91%) updated within 30 days, and none untouched for over a year.
- Three servers are flagged deprecated in the registry and are shown rather than hidden, because a deprecated server still appears in search results elsewhere.
Every figure is reproducible from the tables below. Methodology and citation line at the bottom.
What is an MCP server, and what does this list cover?
An MCP server is a small program that wraps a system, a database, a GitHub repo, a Slack workspace, a browser, and exposes it to an AI client through three primitives: tools the AI can call, resources it can read, and prompts it can reuse. The client talks to it over JSON-RPC, either locally over stdio or remotely over Streamable HTTP. The point of the MCP protocol is that it lets you connect AI securely to a system once and reuse that connection from any client, rather than teaching each LLM about each service.
Inclusion here is automatic. A server is ranked when it appears in the official MCP Registry with a description and a working install method (an npm, PyPI, OCI or NuGet package, or a public remote URL), and clears an adoption floor. Servers marked deleted are excluded; servers marked deprecated are kept and badged, because you deserve to know a thing is deprecated rather than simply not find it.
Most of this is open-source MCP work, built by people who wanted their own AI assistant to reach one more system. One thing this list does that a GitHub “awesome” file cannot: it tells you whether the vendor built the server themselves. That distinction matters more than any score, and it gets its own section below.
How do MCP servers work?
Three roles, and mixing them up is the usual source of confusion. The host is the AI application you already use, Claude Desktop or Cursor or VS Code.
Inside it runs an MCP client, one per connection. The MCP server is the separate program on the other end of that connection, wrapping the thing you want the model to reach.
The client and server speak JSON-RPC over one of two transports. Local servers run over stdio, meaning the host launches the server as a subprocess and talks to it through standard input and output.
Remote servers run over Streamable HTTP against a URL. In this build 162 of 250 servers are stdio and 109 support HTTP.
Once connected, the server advertises what it can do through three primitives. Tools are functions the AI model can call, a SQL query, a pull-request comment, a file write. Resources are data it can read, a file tree or a schema. Prompts are reusable templates the server ships, so a good prompt for the job arrives with the server instead of being something you have to work out.
The sequence is worth internalising because it explains the security model. The client asks the server what it offers, the model decides in natural language which tool to call, the client executes that call, and the result comes back into the model’s context.
The model never touches your system directly. It asks a program you installed to do it, which is why what that program is allowed to do matters more than which model you are running.
What do people use MCP servers for?
The category table further down shows where the supply is. What the supply is for clusters into five jobs, and knowing which one you have is the fastest way to narrow 250 servers to three.
AI coding. The largest use by a wide margin. A developer MCP server lets an assistant interact with repositories, read issues, open pull requests and run builds, so the model works against your actual codebase rather than a pasted snippet. This is what Developer Tools covers, at 81 servers.
Querying live data. A database or analytics server turns a natural-language question into SQL against your real schema, and a cloud platform server does the same for infrastructure state. The value is that the answer is real-time rather than whatever was true when the model was trained. Each connected data source is one more thing the model can check instead of guess.
Giving an assistant memory. Vector stores and knowledge-graph servers act as a persistent knowledge base across sessions, which is what the AI and memory category exists for. It is also the category where the most interesting community work is happening.
Automation across tools. An AI agent with three or four servers connected can run a workflow end to end: read a ticket, query the database, post the summary to Slack. Managing MCP servers becomes a real task at that point, and multiple AI clients sharing one set of them is the case the remote endpoints exist for. This is where MCP stops being a convenience and starts being infrastructure, and it is also where the credential scoping in the vetting section matters most.
Reaching the open web. Search, fetch and browser servers extend AI capabilities past the training cutoff, which is the single most common reason people install their first server.
If none of those is your job, the honest answer is that you may not need an MCP server yet. The protocol is plumbing, and plumbing is only worth installing when there is something to connect.
Is an MCP Server Similar to a REST API?
It is the question that comes up most, and the honest answer is that an MCP server is not a REST API with extra steps: the two operate at different layers. An API is the outlet. MCP is the standard plug.
A REST API is service-specific. Stripe’s API and GitHub’s API are both HTTP, both JSON, and entirely different shapes, so every AI integration has to be written once per service and rewritten whenever either side changes.
An MCP server wraps a system, often by calling that same REST API underneath, and exposes it through one protocol every MCP client already speaks. Write the adapter once, and Claude, Cursor, VS Code and ChatGPT all get the integration.
Three differences that matter in practice:
- Discovery is built in. An MCP client asks a server what tools it has at runtime. A REST client has to be told in advance, usually by a developer reading documentation.
- The description is written for a model, not a developer. Each tool ships a natural-language description the LLM uses to decide when to call it. An OpenAPI spec describes shape, not intent.
- It is stateful and bidirectional. A server can hold a session, stream results, and ask the client for input. REST is request and response.
MCP is not a replacement for your APIs, and it is not a microservice architecture. It is a thin, standardised layer that sits in front of whatever you already have so an AI agent can use it without integration code written for that one service.
If you have one AI client and one service to connect, a direct API call is simpler. The protocol earns its keep at the point where you have several of each.
The Most-Downloaded MCP Servers
There is no universal install counter for the Model Context Protocol, so weekly npm and PyPI download counts are the closest available proxy for a server actually being pulled into real projects. They measure installs rather than bookmarks, which is why they sit alongside stars in the adoption pillar rather than behind them.
Two caveats set the scope of this section, and the second one is the reason our numbers are lower than you will see elsewhere.
Downloads only exist for servers distributed as a package. A remote-only server behind a hosted URL, or one shipped solely as a container image, has nothing for npm or PyPI to count, so it is absent here rather than sitting at zero.
We do not count a download unless the package is the MCP server. Many registry entries point at a general-purpose library that merely also exposes an MCP server, and its download count measures the whole library. The largest single case is browser-use, whose registry entry points at `browser-use`: those 1,721,514 weekly downloads are people installing that library, the overwhelming majority of whom never touch its MCP server. Counting it would be the download equivalent of crediting a server with its parent project’s GitHub stars, which is why the engine already rejects borrowed stars and why we reject borrowed downloads on the same principle.
So 26 servers are excluded from every download total in this report, setting aside 4,015,103 weekly downloads, 51% of the raw sum. They appear in the full list with their download cell marked library-wide rather than as a number that is not comparable to the rest. What remains is 57 servers (23% of the list) whose package is the MCP server itself, drawing 3,897,997 downloads a week between them.
Every download figure and percentage below is computed over that subset. If you see an MCP ranking where a single server holds the overwhelming majority of all downloads, it has almost certainly summed the libraries.
One further limit we cannot correct: a few vendors ship many servers through a single shared package, so that package’s downloads are attributed to the one server the registry attaches it to rather than split across the family.
| Rank | MCP server | Category | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|---|---|
| 1 | Chrome DevTools MCP | Developer Tools | 1,051,748 | First-party | 0d ago | 93/100 |
| 2 | Context7 | Productivity | 871,893 | First-party | 1d ago | 94/100 |
| 3 | Amazon ECS MCP Server | Cloud | 286,535 | First-party | 213d ago | 71/100 |
| 4 | Amazon EKS MCP Server | Cloud | 286,535 | First-party | 225d ago | 71/100 |
| 5 | AWS MCP Server | Cloud | 286,535 | Community | 211d ago | 68/100 |
| 6 | AWS MCP Server | Cloud | 286,535 | Community | 280d ago | 68/100 |
| 7 | Supabase | Databases | 97,145 | First-party | 0d ago | 90/100 |
| 8 | ha-mcp | AI & Memory | 90,471 | First-party | 0d ago | 87/100 |
| 9 | gitlab-mcp | Developer Tools | 65,694 | Community | 2d ago | 85/100 |
| 10 | sentry-mcp | DevOps & Monitoring | 59,130 | First-party | 0d ago | 85/100 |
| 11 | ClickHouse | Databases | 58,094 | First-party | 1d ago | 89/100 |
| 12 | nx-console | AI & Memory | 55,721 | First-party | 0d ago | 88/100 |
| 13 | nx-mcp | AI & Memory | 55,721 | First-party | 0d ago | 88/100 |
| 14 | mongodb-mcp-server | Databases | 54,247 | First-party | 0d ago | 88/100 |
| 15 | XcodeBuildMCP | Productivity | 50,636 | First-party | 0d ago | 90/100 |
Read download counts for what they are even after that filtering. A server pulled into a CI pipeline or rebuilt into a container image every week racks up downloads fast, which is a genuine signal of production use but not the same as a headcount of humans. Chrome DevTools MCP at 1,051,748 a week is well clear of the rest of the field, and servers wired into automated environments get re-installed on every run.
The median packaged server gets 4,078 downloads a week. So the distance between the leader and the middle of the pack is several orders of magnitude, and most MCP servers are still early.
The best MCP servers by quality score
| Rank | MCP server | Category | Quality score | Weekly downloads | First-party | Security |
|---|---|---|---|---|---|---|
| 1 | Repowise | Productivity | 96/100 | not reported | First-party | No known advisory |
| 2 | Context7 | Productivity | 94/100 | 871,893 | First-party | No known advisory |
| 3 | CCXT | Finance | 94/100 | not reported | First-party | No known advisory |
| 4 | mcp-server-browser | Search & Web | 94/100 | not reported | First-party | No known advisory |
| 5 | mcp-server-commands | Developer Tools | 94/100 | not reported | First-party | No known advisory |
| 6 | mcp-server-filesystem | Productivity | 94/100 | not reported | First-party | No known advisory |
| 7 | mcp-server-search | Search & Web | 94/100 | not reported | First-party | No known advisory |
| 8 | Chrome DevTools MCP | Developer Tools | 93/100 | 1,051,748 | First-party | Advisory, patched |
| 9 | Codebase Memory | AI & Memory | 93/100 | 5,825 | Community | No known advisory |
| 10 | browser-use | Search & Web | 93/100 | library-wide | First-party | Advisory, patched |
| 11 | World Monitor | Other | 93/100 | not reported | First-party | Not checked |
| 12 | Scrapling MCP Server | Cloud | 93/100 | not reported | Community | No known advisory |
| 13 | apify-mcp-server | Other | 93/100 | not reported | First-party | Not checked |
| 14 | playwright-mcp | Search & Web | 92/100 | not reported | First-party | Advisory, patched |
| 15 | Serena MCP: the IDE for your agent | Developer Tools | 92/100 | library-wide | First-party | Advisory, patched |
This ranking weighs adoption alongside maintenance, growth, trust and security, so it surfaces servers that are well run rather than only widely installed.
Quality scores here run from 68 to 96 with a median of 83 and a mean of 83.6. That is a wider spread than the model or extension lists we publish, because MCP is young: the difference between a vendor-maintained server with a licence and a repo, and a weekend project published to the registry, is large and the score reflects it.
How is the quality score calculated?
Five weighted pillars: adoption 35%, maintenance 25%, growth 15%, trust 15%, security 10%.
- Adoption (35%) is a composite proxy, stated plainly. Adoption blends weekly package downloads with GitHub stars, both log-scaled. MCP has no universal usage metric, so this is never presented as an exact install count.
- Maintenance (25%) is how recently the server was updated, plus its release cadence.
- Growth (15%) is the 30-day trend, computed from our own stored snapshots, so it appears only once enough history exists. A server with no history shows no trend rather than a fabricated number.
- Trust (15%) is structural: a first-party namespace, an open-source licence, and a public repository.
- Security (10%) maps each server’s npm or PyPI package to the GitHub Advisory Database. A server with no queryable package, remote-only or container-only, shows “not checked” rather than a false all-clear.
No server pays to be listed or ranked, and there are no affiliate links anywhere on this page.
First-Party or community? The signal that matters most
168 of 250 servers (67%) are first-party: published under a reverse-DNS or GitHub-organisation namespace the registry verified as belonging to the vendor. The other 82 (33%) are community builds.
This is the single most decision-relevant fact in the dataset, and it is the one a star count cannot tell you.
An MCP server usually needs credentials. A database server wants a connection string, a Slack server wants a workspace token, a cloud server wants API keys with real permissions.
When the server is first-party, you are trusting the same company you already trusted with the underlying product. When it is community, you are extending that trust to a third party, which may be entirely fine and may be better maintained, but it is a separate decision you should make knowingly.
First-party servers take 82.1% of all download volume, so the ecosystem is consolidating around vendor-built bridges faster than the raw server counts suggest.
Which categories have the most MCP servers?
Developer Tools leads with 81 servers (32% of the list), ahead of Search & Web at 50. By download volume the leading category is Developer Tools, which takes 32.3% from just 81 servers.
| Category | Servers | Weekly downloads | Share | First-party | Top server |
|---|---|---|---|---|---|
| Developer Tools | 81 | 1,257,182 | 32.3% | 54 | mcp-server-commands |
| Search & Web | 50 | 27,303 | 0.7% | 28 | mcp-server-browser |
| AI & Memory | 24 | 208,225 | 5.3% | 12 | Codebase Memory |
| Other | 21 | 4,847 | 0.1% | 18 | World Monitor |
| DevOps & Monitoring | 15 | 61,114 | 1.6% | 12 | Netdata |
| Databases | 15 | 220,938 | 5.7% | 13 | Supabase |
| Productivity | 13 | 922,654 | 23.7% | 8 | Repowise |
| Communication | 10 | none attributable | n/a | 10 | emailmd |
| Cloud | 8 | 1,195,703 | 30.7% | 5 | Scrapling MCP Server |
| Finance | 7 | none attributable | n/a | 4 | CCXT |
| Design | 6 | 31 | <0.1% | 4 | mcp-server-chart |
| All 11 categories | 250 | 3,897,997 | 100% | 168 | Repowise |
That developer tools lead on count is unsurprising: the first people to wire an AI into a system are the people who build systems, and coding assistants were the first clients to ship MCP support. The categories to watch are the sparse ones. A category with two or three servers and real download volume is where the protocol is being adopted faster than it is being served.
How do you install an MCP server?
Every server in this list ships copy-paste configuration for four clients, which is the practical reason to use this page over a plain directory.
| AI client | Servers with a ready install snippet |
|---|---|
| Claude Desktop | 250 of 250 |
| Cursor | 250 of 250 |
| VS Code | 250 of 250 |
| Claude Code | 250 of 250 |
An MCP server runs as a small program, locally over stdio or remotely over Streamable HTTP, that an MCP client connects to. To add one, open your client config and point it at the server: in Claude Desktop edit claude_desktop_config.json, in Cursor edit ~/.cursor/mcp.json, in VS Code add a .vscode/mcp.json, or run “claude mcp add” in Claude Code. Use the MCP Inspector to test a server before wiring it into your assistant.
The transport matters for how you deploy. 162 of 250 servers run over stdio, meaning the client starts them as a local subprocess on your machine. Only 109 support Streamable HTTP, and 93 publish a hosted remote endpoint you can point at without installing anything.
That imbalance is the single biggest practical limitation of MCP today. A local MCP server is simple and private, but it cannot be shared across a team, it does not work from a phone or a browser-only client, and every machine needs its own install and its own credentials.
A hosted MCP server, or an MCP gateway sitting in front of several, is what you deploy when more than one person needs the same connection. The remote MCP endpoints in this build are the ones you can actually centralise:
| Server | Transports | Remote endpoint | Package | Weekly downloads |
|---|---|---|---|---|
| Context7 | stdio, streamable-http | Yes | mcpb, npm | 871,893 |
| Supabase | stdio, streamable-http | Yes | npm | 97,145 |
| CloudBase | stdio, streamable-http | Yes | npm | 20,898 |
| mcp | stdio, streamable-http | Yes | npm | 22 |
| World Monitor | streamable-http | Yes | remote only | not reported |
| Netdata | streamable-http | Yes | remote only | not reported |
| tldraw | streamable-http, sse | Yes | remote only | not reported |
| Metabase | streamable-http | Yes | remote only | not reported |
| mcp-server | streamable-http | Yes | remote only | not reported |
| Reactive Resume | streamable-http | Yes | remote only | not reported |
| PostHog MCP Server | streamable-http, sse | Yes | remote only | not reported |
| GitHub | stdio, streamable-http | Yes | oci | not reported |
Packaging follows the same practical split:
| Package registry | Servers |
|---|---|
| npm | 99 |
| pypi | 58 |
| mcpb | 42 |
| oci | 29 |
| nuget | 4 |
| cargo | 1 |
Which AI clients support MCP, and does ChatGPT?
Yes, ChatGPT supports MCP. OpenAI adopted the protocol in 2025, so it now sits alongside Anthropic’s own clients rather than competing with a rival standard, which is most of why MCP became the default way to connect an AI application to anything.
The clients worth knowing, and what each is good for:
- Claude Desktop and Claude Code are the reference implementations, and every server here ships a copy-paste snippet for both. Claude Code takes a one-line `claude mcp add`, which is the fastest route from this page to a working connection.
- Cursor and VS Code cover the AI coding case, where developer-tool and database servers do the most work. Developer Tools is the largest category in this build at 81 servers, and it is the one most people install first.
- ChatGPT supports MCP through connectors, weighted toward remote HTTP servers rather than local stdio ones. That matters here, because only 93 of 250 servers publish a hosted remote endpoint.
- Gemini CLI, Windsurf and Zed have all added support, and the protocol is open, so the list keeps growing.
One recurring confusion is worth clearing up: GitHub is not itself an MCP server. It is a service, and there is an official GitHub MCP server that wraps it, built by GitHub, which is exactly the first-party distinction the next section is about.
The same is true of Slack, Notion, Supabase and Stripe. The official GitHub MCP server and the Context7 MCP server are the two most commonly recommended starting points for AI coding, and both appear in the full list below with their own figures rather than on somebody’s recommendation.
If your AI client is not on that list, check whether it supports MCP before you pick a server from this page. The snippet coverage column tells you what we generate; it does not tell you what your client accepts.
How do you vet an MCP server before you trust it?
An MCP server is a program you install that runs with your permissions and holds your credentials. The advisory data in the next section says the ecosystem has no disclosed vulnerability problem yet, and that is genuinely different from saying a given server is safe to run.
Four failure modes are specific to this protocol and none of them show up as a CVE:
- Arbitrary code execution. A stdio server is a local subprocess with your user’s access to files, network and shell. There is no sandbox unless you add one.
- Prompt injection. A server returns text into the model’s context. Content it fetched from a web page or an issue comment can carry instructions, and the model may act on them using the other tools you enabled.
- Over-broad credentials. The usual mistake is handing a database server an admin role because it was the key already in the environment.
- Silent tool changes. A server can change the tools it advertises on any update. What you audited in January is not necessarily what runs in March.
The checks that reduce real risk, in the order I would do them:
- Start from the registry entry rather than a search result, and confirm the reverse-DNS namespace matches the vendor you think you are installing. That verification is what the first-party badge below is built on.
- Read the tool list before you enable it. A server publishes the actions it can take, and that list is the real permission surface, not the README.
- Pin a version. An unpinned `npx` or `uvx` command pulls the latest release every launch.
- Run it through the MCP Inspector first and call the tools by hand, before wiring it into an agent that will call them autonomously.
- Give it the narrowest credential that works, read-only by default, and treat write access as a deliberate decision rather than the starting point.
None of this is exotic supply-chain practice. It is the same care you would take with any dependency, applied to one that an LLM gets to drive.
Are MCP Servers Secure?
There are 52 security advisories on record across 17 of the 250 servers here. 126 (50%) have no known advisory and 107 are not checkable because they ship no queryable package.
Of those, two carry an advisory that is still unpatched at the time of this build:
- Desktop Commander (`io.github.wonderwhy-er/desktop-commander`), one open advisory, worst scored CVSS 6.3 (low): DesktopCommanderMCP is vulnerable to SSRF. GHSA-5xx3-j724-wmx5.
- PraisonAI (`io.github.MervinPraison/praisonai`), two open advisories, worst scored CVSS 8.8 (high): Duplicate Advisory: PraisonAI has Memory State Leakage and Path Traversal in MultiAgent Context Handling. GHSA-x44p-gg67-52fc, GHSA-fwh2-95jw-g4j6.
That naming is generated from the advisory data on every refresh rather than written by hand, so a server drops off this list the moment its maintainer ships a fix. Check the linked repository before you act on it: an advisory can be days from a patch, or it can be a duplicate of one already resolved upstream.
The rest is worth stating carefully rather than celebrating. A mostly clean advisory record on an ecosystem this young mostly means nobody has looked yet.
Compare it with our AI WordPress plugins report, where two in three plugins have a disclosed CVE, not because WordPress plugins are worse but because WordPress has a mature vulnerability-reporting culture and a funded advisory feed. MCP has neither yet.
So for most of this list the useful security question is not “does this server have a CVE”. It is what the server can do once you give it credentials:
- Scope the credentials, not the server. Give a database server a read-only role. Give a cloud server a role with the two permissions it needs.
- Prefer first-party for anything holding production keys. 67% of this list qualifies.
- Read the tool list before you enable it. A server advertises the actions it can take; that list is the actual permission surface.
- Test with the MCP Inspector first, before wiring a server into an assistant that will call it autonomously.
- Treat write access as a deliberate choice. Most useful MCP work is read-only.
Are MCP servers actively maintained?
Yes, more so than any other ecosystem we track. The median server was updated 1 days ago, 227 of 250 (91%) within the last 30 days, 242 within 90, and none have gone more than a year. The oldest is 280 days.
That is what a young ecosystem looks like, and it is a genuine advantage of picking an MCP server today: almost nothing here is abandoned. It also means the picture changes fast, which is why this report carries a pull date rather than pretending to be evergreen.
three servers carry a deprecated flag from the registry. They are still listed, with the flag visible, because a deprecated server does not vanish from the rest of the internet just because a directory hides it.
Best MCP Servers by Category
Nobody needs “the best MCP server”. They need one for Postgres, or Slack, or a browser. Each group below is the top servers in that category by quality score, with weekly downloads, first-party status and update recency alongside so you can overrule the ordering on whichever signal you trust most.
Best Developer Tools MCP servers
Developer-tool MCP servers are the largest and most active category, wiring AI into code, repos, and issue trackers.
These servers connect assistants like Claude and Cursor to GitHub, GitLab, Jira, Linear, and your local toolchain so an AI can read code, open pull requests, and triage issues. It is the most mature MCP category and the one most clients ship support for first.
81 servers in this category, 1,257,182 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| mcp-server-commands | not reported | First-party | 0d ago | 94/100 |
| Chrome DevTools MCP | 1,051,748 | First-party | 0d ago | 93/100 |
| Serena MCP: the IDE for your agent | library-wide | First-party | 3d ago | 92/100 |
| firebase-mcp | library-wide | First-party | 0d ago | 92/100 |
| kaneo | 97 | First-party | 0d ago | 91/100 |
| Reactive Resume | not reported | First-party | 0d ago | 91/100 |
Best Search & Web MCP servers
Search and web MCP servers give an AI fresh, real-world context beyond its training cutoff.
Web search, fetch, and browser-automation servers let assistants look things up, read pages, and drive a browser. They are the backbone of grounded, up-to-date AI answers.
50 servers in this category, 27,303 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| mcp-server-browser | not reported | First-party | 0d ago | 94/100 |
| mcp-server-search | not reported | First-party | 0d ago | 94/100 |
| browser-use | library-wide | First-party | 1d ago | 93/100 |
| playwright-mcp | not reported | First-party | 1d ago | 92/100 |
| screenpipe | 2,552 | First-party | 0d ago | 91/100 |
| Skyvern | not reported | First-party | 0d ago | 91/100 |
Best AI & Memory MCP servers
AI and memory MCP servers give assistants persistent memory and a knowledge base.
Vector stores, knowledge graphs, and note systems let an AI remember context across sessions and ground answers in your own documents. The foundation of RAG-style workflows over MCP.
24 servers in this category, 208,225 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| Codebase Memory | 5,825 | Community | 1d ago | 93/100 |
| nx-console | 55,721 | First-party | 0d ago | 88/100 |
| nx-mcp | 55,721 | First-party | 0d ago | 88/100 |
| mcp | 22 | First-party | 1d ago | 88/100 |
| Honcho | not reported | First-party | 1d ago | 88/100 |
| basic-memory | library-wide | First-party | 1d ago | 88/100 |
Best Other MCP servers
Servers across every other niche are joining the MCP ecosystem.
From media and IoT to specialized internal tools, servers of every kind are adopting MCP. This catch-all category captures the breadth of the protocol beyond the obvious developer and data tools.
21 servers in this category, 4,847 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| World Monitor | not reported | First-party | 0d ago | 93/100 |
| apify-mcp-server | not reported | First-party | 1d ago | 93/100 |
| Oh My Posh Validator | not reported | First-party | 2d ago | 91/100 |
| mcp | library-wide | First-party | 23d ago | 91/100 |
| Microsoft Fabric MCP Server | 769 | First-party | 0d ago | 89/100 |
| openstatus | not reported | First-party | 0d ago | 88/100 |
Best DevOps & Monitoring MCP servers
DevOps and monitoring MCP servers expose infrastructure and observability data to AI.
Kubernetes, Docker, Terraform, Sentry, Grafana, and incident tools let an assistant inspect deployments, read logs, and surface errors. Powerful for on-call work, where read-only scopes are the safe default.
15 servers in this category, 61,114 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| Netdata | not reported | First-party | 0d ago | 91/100 |
| kubernetes-mcp-server | 1,984 | First-party | 1d ago | 86/100 |
| mockserver | not reported | First-party | 0d ago | 86/100 |
| sentry-mcp | 59,130 | First-party | 0d ago | 85/100 |
| kubefwd | not reported | First-party | 3d ago | 85/100 |
| mcp-grafana | not reported | First-party | 0d ago | 85/100 |
Best Databases MCP servers
Database MCP servers let an AI query and inspect your data through a controlled interface.
Postgres, MySQL, SQLite, MongoDB, Redis, and managed platforms like Supabase expose schema and read (sometimes write) access over MCP. Useful for natural-language analytics, but scope the credentials carefully, since you are handing an AI a database connection.
15 servers in this category, 220,938 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| Supabase | 97,145 | First-party | 0d ago | 90/100 |
| MCP Toolbox for Databases | not reported | First-party | 0d ago | 90/100 |
| MCP Toolbox for Databases | not reported | First-party | 0d ago | 90/100 |
| ClickHouse | 58,094 | First-party | 1d ago | 89/100 |
| mcp | not reported | First-party | 2d ago | 89/100 |
| mongodb-mcp-server | 54,247 | First-party | 0d ago | 88/100 |
Best Productivity MCP servers
Productivity MCP servers connect AI to notes, docs, files, and project tools.
Notion, Google Drive, filesystem, calendar, and task-manager servers let an AI read and update your working documents and to-dos. Among the most-installed servers because they touch everyday workflows.
13 servers in this category, 922,654 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| Repowise | not reported | First-party | 0d ago | 96/100 |
| Context7 | 871,893 | First-party | 1d ago | 94/100 |
| mcp-server-filesystem | not reported | First-party | 0d ago | 94/100 |
| mcp-server | not reported | First-party | 1d ago | 91/100 |
| XcodeBuildMCP | 50,636 | First-party | 0d ago | 90/100 |
| XcodeBuildMCP | not reported | First-party | 0d ago | 90/100 |
Best Communication MCP servers
Communication MCP servers let AI read and send across chat and email.
Slack, Discord, email, and messaging servers let an assistant summarize threads, draft replies, and post updates. Convenient, but write access to your inbox or channels deserves scrutiny.
10 servers in this category, 0 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| emailmd | library-wide | First-party | 16d ago | 86/100 |
| QR Rápido | not reported | First-party | 1d ago | 85/100 |
| draw.io | not reported | First-party | 39d ago | 85/100 |
| OpenOSINT | not reported | First-party | 3d ago | 85/100 |
| Vexa | not reported | First-party | 4d ago | 84/100 |
| bernstein | library-wide | First-party | 0d ago | 84/100 |
Best Cloud MCP servers
Cloud MCP servers expose AWS, Azure, GCP, and edge platforms to AI.
These servers let an assistant query cloud resources, read storage, and manage services. Among the most sensitive in this list, so prefer scoped, read-only credentials and audit every write.
8 servers in this category, 1,195,703 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| Scrapling MCP Server | not reported | Community | 7d ago | 93/100 |
| Azure MCP Server | not reported | First-party | 0d ago | 87/100 |
| mcp | not reported | First-party | 10d ago | 86/100 |
| next-devtools-mcp | 49,563 | First-party | 2d ago | 85/100 |
| Amazon ECS MCP Server | 286,535 | First-party | 213d ago | 71/100 |
| Amazon EKS MCP Server | 286,535 | First-party | 225d ago | 71/100 |
Best Finance MCP servers
Finance MCP servers connect AI to payments, accounting, and market data.
Stripe, accounting, and banking servers let an assistant read transactions, draft invoices, and pull financial data. High-stakes by nature, so keep write access tightly scoped.
7 servers in this category, 0 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| CCXT | not reported | First-party | 0d ago | 94/100 |
| basebalance.cloud - x402 RPC & MCP gateway | not reported | First-party | 6d ago | 87/100 |
| Finance Toolkit | library-wide | Community | 1d ago | 86/100 |
| tradememory-protocol | not reported | First-party | 3d ago | 85/100 |
| mcp | not reported | First-party | 0d ago | 84/100 |
| tradingview-mcp | not reported | Community | 10d ago | 82/100 |
Best Design MCP servers
Design MCP servers connect AI to design tools and image generation.
Figma, asset, and image-generation servers let an AI read design files, generate visuals, or produce screenshots. A smaller but fast-growing category as design tools open up MCP access.
6 servers in this category, 31 combined weekly downloads.
| Server | Weekly downloads | First-party | Updated | Quality |
|---|---|---|---|---|
| mcp-server-chart | not reported | First-party | 15d ago | 89/100 |
| Figma-Context-MCP | not reported | Community | 1d ago | 88/100 |
| Agent-Native Design | not reported | First-party | 0d ago | 83/100 |
| Figma MCP Server | not reported | First-party | 1d ago | 81/100 |
| openflowkit-mcp | 31 | Community | 23d ago | 79/100 |
| TypeUI | not reported | First-party | 69d ago | 77/100 |
How this compares to the awesome MCP servers lists
Most people looking for MCP servers land on one of three kinds of page, and they are good at different jobs. Worth saying plainly which one you want, because it is not always this one.
The awesome MCP servers lists on GitHub, appcypher’s and punkpeye’s among them, are hand-curated markdown files grouped by category. They are broad, they move fast, and contributors add servers the registry has not seen.
What they cannot do is tell you which of two servers in the same category anyone actually uses, whether either is still maintained, or which one the vendor built. Every entry looks identical on the page.
The marketplaces, mcp.so and mcpservers.org, are browsable catalogues with search and categories. Better for discovery than a flat file, and they carry more servers than this report does. They sort by trending or featured rather than by anything you can audit, and featured placement is not always editorial.
This report is the comparison layer. It starts from the official MCP Registry rather than from curation, so inclusion is mechanical, then adds the figures that let you choose between two candidates: weekly downloads, GitHub stars, last commit, licence, first-party status and known advisories.
That is why it ranks 250 servers and not thousands. Everything below the adoption floor is real but unmeasurable, and a ranking of unmeasurable things is just a list again.
Use an awesome list to find MCP servers you did not know existed. Use this page to pick the right MCP server out of the ones you found.
Neither substitutes for reading the MCP server use cases above and deciding what you are connecting. If you want the raw index with nothing on top, that is the registry itself at registry.modelcontextprotocol.io.
The full list: All 250 MCP servers ranked
Every ranked server, ordered by quality score, with the inputs shown so you can check the arithmetic. Click any server to open its repository. Where a value reads “not reported”, the upstream source did not return it and we would rather say so than print a zero.
| # | MCP server | Category | Weekly downloads | GitHub stars | First-party | Licence | Updated | Security | Quality |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Repowise | Productivity | not reported | 6,403 | First-party | AGPL-3.0 | 0d ago | No known advisory | 96 |
| 2 | Context7 | Productivity | 871,893 | 61,869 | First-party | MIT | 1d ago | No known advisory | 94 |
| 3 | CCXT | Finance | not reported | 43,943 | First-party | MIT | 0d ago | No known advisory | 94 |
| 4 | mcp-server-browser | Search & Web | not reported | 38,932 | First-party | Apache-2.0 | 0d ago | No known advisory | 94 |
| 5 | mcp-server-commands | Developer Tools | not reported | 38,932 | First-party | Apache-2.0 | 0d ago | No known advisory | 94 |
| 6 | mcp-server-filesystem | Productivity | not reported | 38,932 | First-party | Apache-2.0 | 0d ago | No known advisory | 94 |
| 7 | mcp-server-search | Search & Web | not reported | 38,932 | First-party | Apache-2.0 | 0d ago | No known advisory | 94 |
| 8 | Chrome DevTools MCP | Developer Tools | 1,051,748 | 51,622 | First-party | Apache-2.0 | 0d ago | Advisory, patched | 93 |
| 9 | Codebase Memory | AI & Memory | 5,825 | 42,924 | Community | MIT | 1d ago | No known advisory | 93 |
| 10 | browser-use | Search & Web | library-wide | 114,148 | First-party | MIT | 1d ago | Advisory, patched | 93 |
| 11 | World Monitor | Other | not reported | 86,024 | First-party | AGPL-3.0 | 0d ago | Not checked | 93 |
| 12 | Scrapling MCP Server | Cloud | not reported | 80,159 | Community | BSD-3-Clause | 7d ago | No known advisory | 93 |
| 13 | apify-mcp-server | Other | not reported | 6,725 | First-party | MIT | 1d ago | Not checked | 93 |
| 14 | playwright-mcp | Search & Web | not reported | 37,003 | First-party | Apache-2.0 | 1d ago | Advisory, patched | 92 |
| 15 | Serena MCP: the IDE for your agent | Developer Tools | library-wide | 29,170 | First-party | MIT | 3d ago | Advisory, patched | 92 |
| 16 | firebase-mcp | Developer Tools | library-wide | 4,465 | First-party | MIT | 0d ago | Advisory, patched | 92 |
| 17 | screenpipe | Search & Web | 2,552 | 21,530 | First-party | not reported | 0d ago | No known advisory | 91 |
| 18 | kaneo | Developer Tools | 97 | 9,041 | First-party | MIT | 0d ago | No known advisory | 91 |
| 19 | Netdata | DevOps & Monitoring | not reported | 80,483 | First-party | GPL-3.0 | 0d ago | Not checked | 91 |
| 20 | mcp-server | Productivity | not reported | 43,506 | First-party | AGPL-3.0 | 1d ago | Not checked | 91 |
| 21 | Reactive Resume | Developer Tools | not reported | 42,462 | First-party | MIT | 0d ago | Not checked | 91 |
| 22 | GitHub | Developer Tools | not reported | 32,860 | First-party | MIT | 1d ago | Not checked | 91 |
| 23 | Oh My Posh Validator | Other | not reported | 23,443 | First-party | MIT | 2d ago | Not checked | 91 |
| 24 | Skyvern | Search & Web | not reported | 22,971 | First-party | AGPL-3.0 | 0d ago | Not checked | 91 |
| 25 | FunASR | Developer Tools | not reported | 20,281 | First-party | MIT | 1d ago | Not checked | 91 |
| 26 | Compiler Explorer | Developer Tools | not reported | 19,059 | First-party | BSD-2-Clause | 1d ago | Not checked | 91 |
| 27 | Keploy | Developer Tools | not reported | 18,452 | First-party | Apache-2.0 | 0d ago | Not checked | 91 |
| 28 | Firecrawl MCP Server | Search & Web | not reported | 7,434 | First-party | MIT | 1d ago | No known advisory | 91 |
| 29 | mcp | Other | library-wide | 55 | First-party | Apache-2.0 | 23d ago | Advisory, patched | 91 |
| 30 | Supabase | Databases | 97,145 | 2,901 | First-party | Apache-2.0 | 0d ago | No known advisory | 90 |
| 31 | XcodeBuildMCP | Productivity | 50,636 | 6,369 | First-party | MIT | 0d ago | No known advisory | 90 |
| 32 | dbx | Developer Tools | 4,735 | 19,069 | Community | Apache-2.0 | 0d ago | No known advisory | 90 |
| 33 | claude-flow | Developer Tools | library-wide | 72,049 | Community | MIT | 0d ago | No known advisory | 90 |
| 34 | Agent Skills Search Server | Search & Web | not reported | 25,224 | First-party | Apache-2.0 | 33d ago | Not checked | 90 |
| 35 | MCP Toolbox for Databases | Databases | not reported | 16,366 | First-party | Apache-2.0 | 0d ago | Not checked | 90 |
| 36 | MCP Toolbox for Databases | Databases | not reported | 16,366 | First-party | Apache-2.0 | 0d ago | Not checked | 90 |
| 37 | OpenMetadata | Search & Web | not reported | 15,169 | First-party | Apache-2.0 | 0d ago | Not checked | 90 |
| 38 | Coder | Developer Tools | not reported | 14,437 | First-party | AGPL-3.0 | 0d ago | Not checked | 90 |
| 39 | XcodeBuildMCP | Productivity | not reported | 6,369 | First-party | MIT | 0d ago | No known advisory | 90 |
| 40 | ClickHouse | Databases | 58,094 | 869 | First-party | Apache-2.0 | 1d ago | No known advisory | 89 |
| 41 | Windows-MCP | Developer Tools | 13,992 | 6,971 | First-party | MIT | 1d ago | Advisory, patched | 89 |
| 42 | Microsoft Fabric MCP Server | Other | 769 | 3,664 | First-party | MIT | 0d ago | No known advisory | 89 |
| 43 | mcp | Databases | not reported | 12,076 | First-party | Apache-2.0 | 2d ago | Not checked | 89 |
| 44 | agent-device | Developer Tools | library-wide | 4,520 | First-party | MIT | 0d ago | No known advisory | 89 |
| 45 | agent-device | Developer Tools | not reported | 4,520 | First-party | MIT | 0d ago | No known advisory | 89 |
| 46 | mcp-server-chart | Design | not reported | 4,363 | First-party | MIT | 15d ago | No known advisory | 89 |
| 47 | nx-console | AI & Memory | 55,721 | 1,414 | First-party | MIT | 0d ago | No known advisory | 88 |
| 48 | nx-mcp | AI & Memory | 55,721 | 1,414 | First-party | MIT | 0d ago | No known advisory | 88 |
| 49 | mongodb-mcp-server | Databases | 54,247 | 1,126 | First-party | Apache-2.0 | 0d ago | No known advisory | 88 |
| 50 | mobile-mcp | Developer Tools | 14,384 | 6,644 | First-party | Apache-2.0 | 0d ago | Advisory, patched | 88 |
| 51 | mcp | AI & Memory | 22 | 3,434 | First-party | Apache-2.0 | 1d ago | No known advisory | 88 |
| 52 | tldraw | Developer Tools | not reported | 50,279 | First-party | not reported | 0d ago | Not checked | 88 |
| 53 | Metabase | Search & Web | not reported | 49,192 | First-party | not reported | 0d ago | Not checked | 88 |
| 54 | PostHog MCP Server | Developer Tools | not reported | 39,740 | First-party | not reported | 0d ago | Not checked | 88 |
| 55 | Figma-Context-MCP | Design | not reported | 15,837 | Community | MIT | 1d ago | Advisory, patched | 88 |
| 56 | mobilerun | Developer Tools | not reported | 9,346 | First-party | MIT | 0d ago | Not checked | 88 |
| 57 | openstatus | Other | not reported | 9,101 | First-party | AGPL-3.0 | 0d ago | Not checked | 88 |
| 58 | Honcho | AI & Memory | not reported | 7,111 | First-party | AGPL-3.0 | 1d ago | Not checked | 88 |
| 59 | basic-memory | AI & Memory | library-wide | 3,930 | First-party | AGPL-3.0 | 1d ago | No known advisory | 88 |
| 60 | DBHub | Databases | not reported | 3,495 | First-party | MIT | 1d ago | No known advisory | 88 |
| 61 | sem | AI & Memory | library-wide | 3,341 | First-party | Apache-2.0 | 0d ago | No known advisory | 88 |
| 62 | BoostedTravel | Search & Web | not reported | 2,001 | First-party | not reported | 0d ago | No known advisory | 88 |
| 63 | ha-mcp | AI & Memory | 90,471 | 4,687 | First-party | MIT | 0d ago | Advisory, patched | 87 |
| 64 | Zotero MCP | Search & Web | 2,818 | 4,990 | Community | MIT | 17d ago | No known advisory | 87 |
| 65 | basebalance.cloud - x402 RPC & MCP gateway | Finance | not reported | 6,308 | First-party | MIT | 6d ago | Not checked | 87 |
| 66 | exa | Search & Web | not reported | 4,990 | First-party | MIT | 21d ago | Not checked | 87 |
| 67 | Azure MCP Server | Cloud | not reported | 3,664 | First-party | MIT | 0d ago | Advisory, patched | 87 |
| 68 | Argent | Search & Web | library-wide | 2,772 | First-party | Apache-2.0 | 0d ago | No known advisory | 87 |
| 69 | Semiotic | Developer Tools | library-wide | 2,705 | First-party | Apache-2.0 | 0d ago | No known advisory | 87 |
| 70 | brightdata-mcp | Search & Web | not reported | 2,635 | First-party | MIT | 30d ago | No known advisory | 87 |
| 71 | tavily-mcp | Search & Web | not reported | 2,378 | First-party | MIT | 1d ago | No known advisory | 87 |
| 72 | mcp-server | Developer Tools | 48,546 | 99 | First-party | Apache-2.0 | 0d ago | No known advisory | 86 |
| 73 | CloudBase | Developer Tools | 20,898 | 1,102 | First-party | MIT | 0d ago | No known advisory | 86 |
| 74 | Power BI Modeling MCP Server | Other | 4,078 | 1,142 | First-party | MIT | 3d ago | No known advisory | 86 |
| 75 | kubernetes-mcp-server | DevOps & Monitoring | 1,984 | 2,077 | First-party | Apache-2.0 | 1d ago | No known advisory | 86 |
| 76 | Claude Code Ultimate Guide | Search & Web | 427 | 5,950 | Community | CC-BY-SA-4.0 | 1d ago | No known advisory | 86 |
| 77 | Finance Toolkit | Finance | library-wide | 5,319 | Community | MIT | 1d ago | No known advisory | 86 |
| 78 | Jitsu | Other | not reported | 5,068 | First-party | MIT | 0d ago | Not checked | 86 |
| 79 | mockserver | DevOps & Monitoring | not reported | 4,969 | First-party | Apache-2.0 | 0d ago | Not checked | 86 |
| 80 | mcp | Cloud | not reported | 4,183 | First-party | Apache-2.0 | 10d ago | Not checked | 86 |
| 81 | PaperBanana | Developer Tools | not reported | 2,347 | First-party | MIT | 2d ago | No known advisory | 86 |
| 82 | pg-aiguide | Developer Tools | not reported | 1,835 | First-party | Apache-2.0 | 2d ago | No known advisory | 86 |
| 83 | emailmd | Communication | library-wide | 1,371 | First-party | MIT | 16d ago | No known advisory | 86 |
| 84 | gitlab-mcp | Developer Tools | 65,694 | 1,967 | Community | MIT | 2d ago | No known advisory | 85 |
| 85 | sentry-mcp | DevOps & Monitoring | 59,130 | 846 | First-party | not reported | 0d ago | No known advisory | 85 |
| 86 | next-devtools-mcp | Cloud | 49,563 | 819 | First-party | not reported | 2d ago | No known advisory | 85 |
| 87 | brave-search-mcp-server | Search & Web | 7,679 | 1,433 | First-party | MIT | 1d ago | No known advisory | 85 |
| 88 | mcp-server-browserbase | Search & Web | 1,838 | 3,404 | First-party | Apache-2.0 | 53d ago | No known advisory | 85 |
| 89 | IWE | AI & Memory | 465 | 1,634 | First-party | Apache-2.0 | 2d ago | No known advisory | 85 |
| 90 | Business Contact Finder | Other | not reported | 7,235 | First-party | not reported | 1d ago | Not checked | 85 |
| 91 | Contractor Licence Changes | Other | not reported | 7,235 | First-party | not reported | 1d ago | Not checked | 85 |
| 92 | Contact Affordance | Other | not reported | 7,235 | First-party | not reported | 1d ago | Not checked | 85 |
| 93 | Domain Liveness | Other | not reported | 7,235 | First-party | not reported | 1d ago | Not checked | 85 |
| 94 | QR Rápido | Communication | not reported | 7,235 | First-party | not reported | 1d ago | Not checked | 85 |
| 95 | draw.io | Communication | not reported | 5,383 | First-party | Apache-2.0 | 39d ago | Not checked | 85 |
| 96 | kubefwd | DevOps & Monitoring | not reported | 4,167 | First-party | Apache-2.0 | 3d ago | Not checked | 85 |
| 97 | AdsAgent | Developer Tools | not reported | 3,723 | First-party | MIT | 0d ago | Not checked | 85 |
| 98 | NotFair | Search & Web | not reported | 3,723 | First-party | MIT | 0d ago | Not checked | 85 |
| 99 | NotFair-MetaAds | Developer Tools | not reported | 3,723 | First-party | MIT | 0d ago | Not checked | 85 |
| 100 | mcp-grafana | DevOps & Monitoring | not reported | 3,449 | First-party | Apache-2.0 | 0d ago | Not checked | 85 |
| 101 | Radar Kubernetes MCP Server | DevOps & Monitoring | not reported | 3,322 | First-party | Apache-2.0 | 0d ago | Not checked | 85 |
| 102 | mcp | Other | not reported | 2,671 | First-party | Apache-2.0 | 0d ago | Not checked | 85 |
| 103 | tooluniverse | Developer Tools | not reported | 1,680 | First-party | Apache-2.0 | 2d ago | No known advisory | 85 |
| 104 | OpenOSINT | Communication | not reported | 1,555 | First-party | MIT | 3d ago | No known advisory | 85 |
| 105 | ros-mcp-server | Developer Tools | not reported | 1,448 | First-party | Apache-2.0 | 1d ago | No known advisory | 85 |
| 106 | tradememory-protocol | Finance | not reported | 1,416 | First-party | MIT | 3d ago | No known advisory | 85 |
| 107 | voicemode | Other | not reported | 1,358 | First-party | MIT | 2d ago | No known advisory | 85 |
| 108 | mcp | Search & Web | not reported | 1,313 | First-party | Apache-2.0 | 0d ago | No known advisory | 85 |
| 109 | arxiv-mcp-server | Search & Web | 9,960 | 3,132 | Community | Apache-2.0 | 15d ago | No known advisory | 84 |
| 110 | mcp-neo4j-data-modeling | Developer Tools | 2,288 | 980 | First-party | MIT | 2d ago | No known advisory | 84 |
| 111 | pyscn | Developer Tools | 1,041 | 1,061 | First-party | MIT | 1d ago | No known advisory | 84 |
| 112 | mcp-neo4j-aura-manager | Databases | 23 | 980 | First-party | MIT | 2d ago | No known advisory | 84 |
| 113 | Ouroboros | Developer Tools | library-wide | 5,807 | Community | MIT | 0d ago | Advisory, patched | 84 |
| 114 | Unity-MCP | DevOps & Monitoring | not reported | 4,238 | Community | Apache-2.0 | 7d ago | Not checked | 84 |
| 115 | socraticode | Search & Web | not reported | 3,294 | Community | AGPL-3.0 | 0d ago | No known advisory | 84 |
| 116 | Vexa | Communication | not reported | 2,775 | First-party | Apache-2.0 | 4d ago | Not checked | 84 |
| 117 | claude-real-video | Search & Web | library-wide | 2,129 | Community | MIT | 0d ago | No known advisory | 84 |
| 118 | mcp | Finance | not reported | 1,798 | First-party | MIT | 0d ago | Not checked | 84 |
| 119 | Persome | AI & Memory | not reported | 1,321 | First-party | Apache-2.0 | 21d ago | No known advisory | 84 |
| 120 | bernstein | Communication | library-wide | 1,157 | First-party | Apache-2.0 | 0d ago | No known advisory | 84 |
| 121 | bernstein | Developer Tools | library-wide | 1,157 | First-party | Apache-2.0 | 0d ago | No known advisory | 84 |
| 122 | mcp-neo4j-memory | AI & Memory | not reported | 980 | First-party | MIT | 2d ago | No known advisory | 84 |
| 123 | Memorix | AI & Memory | not reported | 756 | Community | Apache-2.0 | 0d ago | No known advisory | 84 |
| 124 | StackQL MCP Server | Databases | 121 | 958 | First-party | MIT | 2d ago | No known advisory | 83 |
| 125 | Claude Code Explorer MCP | Search & Web | 28 | 6,244 | Community | not reported | 4d ago | No known advisory | 83 |
| 126 | blitz | Developer Tools | 22 | 1,752 | First-party | Apache-2.0 | 59d ago | No known advisory | 83 |
| 127 | Airweave Search | Search & Web | not reported | 6,564 | First-party | MIT | 98d ago | No known advisory | 83 |
| 128 | Agent-Native Analytics | Developer Tools | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 129 | Agent-Native Assets | Search & Web | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 130 | Agent-Native Brain | Communication | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 131 | Agent-Native Calendar | Productivity | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 132 | Agent-Native Chat | Communication | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 133 | Agent-Native Clips | Developer Tools | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 134 | Agent-Native Content | AI & Memory | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 135 | Agent-Native Design | Design | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 136 | Agent-Native Dispatch | Communication | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 137 | Agent-Native Forms | Developer Tools | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 138 | Agent-Native Mail | Communication | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 139 | Agent-Native Plan | Developer Tools | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 140 | Agent-Native Slides | Developer Tools | not reported | 4,717 | First-party | not reported | 0d ago | Not checked | 83 |
| 141 | EdgarTools | Developer Tools | not reported | 2,697 | Community | MIT | 2d ago | No known advisory | 83 |
| 142 | LetsFG - Flights & Hotels | Search & Web | not reported | 2,001 | First-party | not reported | 0d ago | No known advisory | 83 |
| 143 | Microsoft Learn MCP | Productivity | not reported | 1,884 | First-party | CC-BY-4.0 | 1d ago | Not checked | 83 |
| 144 | Appllama | Other | not reported | 1,451 | First-party | MIT | 5d ago | Not checked | 83 |
| 145 | ArcadeDB MCP Server | Databases | not reported | 1,146 | First-party | Apache-2.0 | 0d ago | Not checked | 83 |
| 146 | GoModel | DevOps & Monitoring | not reported | 1,144 | First-party | MIT | 0d ago | Not checked | 83 |
| 147 | Atlassian Rovo MCP Server | Search & Web | not reported | 1,033 | First-party | Apache-2.0 | 0d ago | Not checked | 83 |
| 148 | Microsoft 365 MCP Server | Developer Tools | not reported | 965 | First-party | MIT | 0d ago | No known advisory | 83 |
| 149 | PDF Reader MCP | Developer Tools | not reported | 921 | First-party | MIT | 1d ago | No known advisory | 83 |
| 150 | UniFi Network MCP | Developer Tools | 18,669 | 805 | Community | MIT | 1d ago | No known advisory | 82 |
| 151 | mcp-neo4j-cypher | Databases | 11,226 | 980 | First-party | MIT | 2d ago | Advisory, patched | 82 |
| 152 | tradingview-mcp | Finance | not reported | 4,419 | Community | MIT | 10d ago | Not checked | 82 |
| 153 | Tolgee | Search & Web | not reported | 4,098 | First-party | not reported | 1d ago | Not checked | 82 |
| 154 | ShipSwift | Developer Tools | not reported | 2,988 | First-party | MIT | 33d ago | Not checked | 82 |
| 155 | Mindwtr | Productivity | not reported | 1,998 | Community | AGPL-3.0 | 0d ago | No known advisory | 82 |
| 156 | TerraVision | DevOps & Monitoring | not reported | 1,629 | Community | AGPL-3.0 | 7d ago | No known advisory | 82 |
| 157 | Microsoft NuGet | AI & Memory | not reported | 1,553 | First-party | not reported | 1d ago | No known advisory | 82 |
| 158 | Terraform | DevOps & Monitoring | not reported | 1,527 | First-party | MPL-2.0 | 0d ago | Not checked | 82 |
| 159 | Superlog | Other | not reported | 1,443 | First-party | Apache-2.0 | 1d ago | Not checked | 82 |
| 160 | Measure | Other | not reported | 1,392 | First-party | Apache-2.0 | 0d ago | Not checked | 82 |
| 161 | monitor | DevOps & Monitoring | not reported | 1,300 | First-party | not reported | 2d ago | No known advisory | 82 |
| 162 | Agent Swarm | Databases | library-wide | 760 | First-party | MIT | 0d ago | No known advisory | 82 |
| 163 | Docling MCP | Developer Tools | not reported | 739 | First-party | MIT | 4d ago | No known advisory | 82 |
| 164 | skylos | Developer Tools | library-wide | 718 | Community | Apache-2.0 | 0d ago | No known advisory | 82 |
| 165 | Nova3D | Developer Tools | not reported | 704 | First-party | MIT | 30d ago | No known advisory | 82 |
| 166 | Emilia Protocol | Developer Tools | not reported | 650 | First-party | Apache-2.0 | 1d ago | No known advisory | 82 |
| 167 | yutu | Developer Tools | not reported | 635 | First-party | Apache-2.0 | 4d ago | No known advisory | 82 |
| 168 | ref-tools-mcp | Search & Web | 334 | 1,173 | First-party | MIT | 77d ago | No known advisory | 81 |
| 169 | Desktop Commander | Productivity | library-wide | 9,538 | Community | MIT | 1d ago | OPEN ADVISORY | 81 |
| 170 | wigolo | Search & Web | not reported | 5,225 | Community | not reported | 0d ago | No known advisory | 81 |
| 171 | RevoGrid DataGrid MCP | AI & Memory | not reported | 3,442 | Community | MIT | 0d ago | Not checked | 81 |
| 172 | RevoGrid DataGrid MCP Pro | Productivity | not reported | 3,442 | Community | MIT | 0d ago | Not checked | 81 |
| 173 | Figma MCP Server | Design | not reported | 1,977 | First-party | not reported | 1d ago | Not checked | 81 |
| 174 | MCP Server for WinDbg Crash Analysis | AI & Memory | not reported | 1,567 | Community | MIT | 1d ago | No known advisory | 81 |
| 175 | minutes | Search & Web | not reported | 1,472 | Community | MIT | 1d ago | No known advisory | 81 |
| 176 | AutoTS | Search & Web | not reported | 1,429 | Community | MIT | 17d ago | No known advisory | 81 |
| 177 | pm-claude-skills | Developer Tools | not reported | 1,356 | Community | MIT | 1d ago | No known advisory | 81 |
| 178 | SafeDep Vet MCP | Developer Tools | not reported | 1,106 | First-party | Apache-2.0 | 8d ago | Not checked | 81 |
| 179 | QueryWeaver | Databases | not reported | 1,075 | First-party | AGPL-3.0 | 3d ago | Not checked | 81 |
| 180 | Scaleway MCP server | Developer Tools | not reported | 997 | First-party | Apache-2.0 | 0d ago | Not checked | 81 |
| 181 | HOL Guard | Developer Tools | not reported | 591 | First-party | Apache-2.0 | 0d ago | No known advisory | 81 |
| 182 | jCodemunch MCP | Developer Tools | 7,646 | 2,679 | Community | not reported | 0d ago | No known advisory | 80 |
| 183 | jshookmcp | Search & Web | 482 | 1,984 | Community | AGPL-3.0 | 1d ago | Advisory, patched | 80 |
| 184 | idea-reality-mcp | Developer Tools | 53 | 815 | First-party | MIT | 31d ago | No known advisory | 80 |
| 185 | Usertour MCP Server | Developer Tools | not reported | 2,302 | First-party | not reported | 0d ago | Not checked | 80 |
| 186 | LLM Sandbox | Developer Tools | not reported | 1,119 | Community | MIT | 1d ago | No known advisory | 80 |
| 187 | unreal-engine-mcp | Developer Tools | not reported | 864 | Community | MIT | 0d ago | No known advisory | 80 |
| 188 | UniFi Protect MCP | Developer Tools | 4,870 | 805 | Community | MIT | 1d ago | No known advisory | 79 |
| 189 | UniFi Access MCP | Developer Tools | 2,185 | 805 | Community | MIT | 1d ago | No known advisory | 79 |
| 190 | reddit-mcp-buddy | Search & Web | 1,185 | 820 | Community | MIT | 25d ago | No known advisory | 79 |
| 191 | agent-harnesses | Developer Tools | 153 | 836 | Community | CC-BY-SA-4.0 | 2d ago | No known advisory | 79 |
| 192 | openflowkit-mcp | Design | 31 | 758 | Community | MIT | 23d ago | No known advisory | 79 |
| 193 | PraisonAI | Developer Tools | library-wide | 9,043 | Community | MIT | 1d ago | OPEN ADVISORY | 79 |
| 194 | strata | Other | not reported | 5,802 | First-party | Apache-2.0 | 102d ago | Not checked | 79 |
| 195 | openshorts | Developer Tools | not reported | 3,986 | Community | not reported | 1d ago | Not checked | 79 |
| 196 | Google Ads + Meta Ads + GA4 MCP | Developer Tools | not reported | 1,715 | Community | MIT | 2d ago | Not checked | 79 |
| 197 | testkube-mcp | DevOps & Monitoring | not reported | 1,656 | First-party | not reported | 1d ago | Not checked | 79 |
| 198 | brave | Search & Web | not reported | 1,433 | Community | MIT | 1d ago | Not checked | 79 |
| 199 | treg.to | Developer Tools | not reported | 1,352 | First-party | not reported | 0d ago | Not checked | 79 |
| 200 | Octocode MCP - AI Context Platform | Search & Web | not reported | 932 | Community | MIT | 1d ago | No known advisory | 79 |
| 201 | projectmem | AI & Memory | library-wide | 804 | Community | MIT | 4d ago | No known advisory | 79 |
| 202 | deja-vu | AI & Memory | library-wide | 802 | Community | MIT | 0d ago | No known advisory | 79 |
| 203 | comfyui-mcp | Developer Tools | not reported | 738 | Community | MIT | 1d ago | No known advisory | 79 |
| 204 | xbbg MCP | Developer Tools | not reported | 716 | First-party | Apache-2.0 | 0d ago | Not checked | 79 |
| 205 | SandBase Harness | Developer Tools | not reported | 642 | First-party | Apache-2.0 | 5d ago | Not checked | 79 |
| 206 | Marmot Data Catalog | Search & Web | not reported | 613 | First-party | MIT | 0d ago | Not checked | 79 |
| 207 | Hex Line | Productivity | 125 | 558 | Community | MIT | 6d ago | No known advisory | 78 |
| 208 | Hex Graph | Developer Tools | 91 | 558 | Community | MIT | 6d ago | No known advisory | 78 |
| 209 | Apple Health | Databases | 82 | 567 | Community | MIT | 15d ago | No known advisory | 78 |
| 210 | Hex Research | Developer Tools | 13 | 558 | Community | MIT | 6d ago | No known advisory | 78 |
| 211 | browserbasehq-mcp-browserbase | Search & Web | not reported | 3,404 | Community | Apache-2.0 | 53d ago | Not checked | 78 |
| 212 | Solana MCP by Vybe | Developer Tools | not reported | 1,223 | First-party | not reported | 18d ago | Not checked | 78 |
| 213 | Solana MCP by Vybe | Developer Tools | not reported | 1,223 | First-party | not reported | 18d ago | Not checked | 78 |
| 214 | SearXNG Search | Search & Web | not reported | 1,220 | Community | MIT | 0d ago | Advisory, patched | 78 |
| 215 | ref-tools-mcp | Search & Web | not reported | 1,173 | First-party | MIT | 77d ago | Not checked | 78 |
| 216 | docs | Search & Web | not reported | 1,156 | First-party | not reported | 0d ago | Not checked | 78 |
| 217 | token-savior | AI & Memory | library-wide | 1,147 | Community | MIT | 32d ago | No known advisory | 78 |
| 218 | SSH - policy-gated remote access | Developer Tools | not reported | 717 | Community | MIT | 1d ago | No known advisory | 78 |
| 219 | MCP Server for Excel | Developer Tools | not reported | 682 | Community | MIT | 0d ago | No known advisory | 78 |
| 220 | obsidian-mcp-server | Search & Web | not reported | 675 | Community | Apache-2.0 | 2d ago | No known advisory | 78 |
| 221 | Vestige | AI & Memory | not reported | 617 | Community | AGPL-3.0 | 0d ago | No known advisory | 78 |
| 222 | haiku.rag | Search & Web | not reported | 606 | Community | MIT | 0d ago | No known advisory | 78 |
| 223 | coolify | Search & Web | not reported | 589 | Community | MIT | 0d ago | No known advisory | 78 |
| 224 | TypeUI | Design | not reported | 1,904 | First-party | not reported | 69d ago | Not checked | 77 |
| 225 | Gearboy MCP Server | Developer Tools | not reported | 1,200 | Community | GPL-3.0 | 2d ago | Not checked | 77 |
| 226 | Magic Cloud | Databases | not reported | 1,189 | Community | MIT | 2d ago | Not checked | 77 |
| 227 | loki-mode | Developer Tools | library-wide | 1,056 | Community | not reported | 0d ago | No known advisory | 77 |
| 228 | ChiR24-unreal_mcp | Other | not reported | 864 | Community | MIT | 0d ago | Not checked | 77 |
| 229 | ChiR24-unreal_mcp_server | AI & Memory | not reported | 864 | Community | MIT | 0d ago | Not checked | 77 |
| 230 | runno | Developer Tools | 35 | 771 | Community | MIT | 41d ago | No known advisory | 76 |
| 231 | token-savior-recall | AI & Memory | library-wide | 1,147 | Community | MIT | 32d ago | No known advisory | 76 |
| 232 | CRW Web Scraper | Search & Web | not reported | 973 | Community | AGPL-3.0 | 0d ago | Not checked | 76 |
| 233 | fastCRW | Search & Web | not reported | 973 | Community | AGPL-3.0 | 0d ago | Not checked | 76 |
| 234 | Tapo MCP | Developer Tools | not reported | 799 | Community | MIT | 1d ago | Not checked | 76 |
| 235 | SonarQube MCP Server | DevOps & Monitoring | not reported | 644 | First-party | not reported | 0d ago | Not checked | 76 |
| 236 | Compartment | AI & Memory | library-wide | 586 | Community | Apache-2.0 | 0d ago | No known advisory | 76 |
| 237 | OpenBrand | Developer Tools | 22 | 791 | First-party | MIT | 122d ago | No known advisory | 75 |
| 238 | ref-tools-ref-tools-mcp | Other | not reported | 1,173 | Community | MIT | 77d ago | Not checked | 75 |
| 239 | FableCut | Search & Web | not reported | 661 | Community | MIT | 0d ago | Not checked | 75 |
| 240 | rustunnel | Developer Tools | not reported | 655 | Community | AGPL-3.0 | 8d ago | Not checked | 75 |
| 241 | TokenSave | Productivity | not reported | 625 | Community | MIT | 3d ago | Not checked | 75 |
| 242 | Sessy - Amazon SES observability | DevOps & Monitoring | not reported | 924 | Community | not reported | 2d ago | Not checked | 73 |
| 243 | TickDB Market Data | Finance | not reported | 779 | Community | MIT | 83d ago | Not checked | 73 |
| 244 | sqz | AI & Memory | not reported | 621 | Community | not reported | 14d ago | Not checked | 72 |
| 245 | Kody | Search & Web | not reported | 610 | Community | not reported | 0d ago | Not checked | 72 |
| 246 | Amazon ECS MCP Server | Cloud | 286,535 | not reported | First-party | not reported | 213d ago | No known advisory | 71 |
| 247 | Amazon EKS MCP Server | Cloud | 286,535 | not reported | First-party | not reported | 225d ago | No known advisory | 71 |
| 248 | AWS MCP Server | Cloud | 286,535 | not reported | Community | not reported | 211d ago | No known advisory | 68 |
| 249 | AWS MCP Server | Cloud | 286,535 | not reported | Community | not reported | 280d ago | No known advisory | 68 |
| 250 | hustcc-mcp-mermaid | Other | not reported | 634 | Community | MIT | 119d ago | Not checked | 68 |
What this data says about MCP right now
Five conclusions, all checkable against the tables above:
- MCP is still mostly a local-process protocol. 162 of 250 servers run over stdio, and while 93 now publish a hosted endpoint, that leaves the majority installable only on the machine that runs the AI client. It is shifting, but MCP is still a developer-machine technology more than a team one.
- First-party is the signal to buy on. 67% of servers are vendor-built and they take 82.1% of all download volume, so the ecosystem is consolidating around official bridges.
- Adoption is concentrated, but less than published figures suggest. Once library-wide packages are set aside, the most-downloaded MCP server takes 27.0% of attributable volume against a median of 4,078 a week. That is still a steep curve, and it is the honest version of it.
- The clean security record is a maturity artefact, not a guarantee. 52 advisories across 250 servers, all patched, is not evidence that MCP servers are safe: 107 of them publish nothing an advisory database can be queried against, so their record is unknown rather than clean. Scope every credential you hand a server as if a disclosure were coming.
- Nothing here is abandoned yet. Median 1 days since update, which is the healthiest maintenance profile of any ecosystem in this series.
“The star counts and the awesome-lists tell you what developers find interesting. What actually decides whether you should install an MCP server is narrower: did the vendor build it, what credentials does it want, and can you run it anywhere other than your own laptop. On that last one the answer is still mostly no, and that is the real state of MCP in 2026, whatever the server count suggests.”
Alston Antony, founder of zplatform.ai and Senior Digital Marketing Manager at Brainstorm Force
Connecting AI to other places instead? The best AI WordPress plugins report carries a full CVE audit, the best Hugging Face models report ranks open-weight models on download data, and the AI Chrome extensions and AI Firefox add-ons reports cover the browser. More sit in best AI tools.
Frequently Asked Questions
What is an MCP server?
An MCP server is a program that exposes tools, data and prompts to AI applications through the Model Context Protocol. It acts as a standardised bridge so an assistant like Claude, Cursor or ChatGPT can read from and act on an external system, a database, a GitHub repo, a Slack workspace, without custom integration code.
What does MCP stand for?
MCP stands for Model Context Protocol, an open standard introduced by Anthropic in November 2024 for connecting AI applications to external tools and data sources.
What are the best MCP servers?
By weekly package downloads the leader is Chrome DevTools MCP at 1,051,748. By our composite quality score, which weighs maintenance, trust and security alongside adoption, the leader is Repowise at 96/100. Both full rankings are above, and the category sections narrow it to the job you actually have.
Where can I find a list of MCP servers?
This page is one, built from the official MCP Registry and enriched with adoption, maintenance and security data. The canonical source is the registry itself at registry.modelcontextprotocol.io; this report adds the comparison layer on top so you can rank and filter rather than only browse.
Are MCP servers free?
Most are free and open-source, and you install them at no cost. Some connect to paid third-party services (a database host, a SaaS API) that need your own account and API key, but the server software itself is typically free. Every row shows the install method and package registry.
What are the most popular MCP servers?
By weekly downloads: Chrome DevTools MCP (1,051,748), Context7 (871,893) and Amazon ECS MCP Server (286,535). Popularity clusters around Developer Tools, which takes 32.3% of all download volume from 81 servers.
MCP versus RAG: what is the difference?
RAG retrieves relevant text and injects it into a prompt. MCP is a connection protocol that lets an AI call live tools and read data sources directly, including taking actions. They are complementary: an MCP server can power retrieval for RAG, but MCP also covers actions RAG alone does not.
What are the best MCP servers for Claude Code?
Every server in this list works with Claude Code and ships a copy-paste “claude mcp add” command. For coding specifically the useful groups are developer tools, filesystem, databases, and browser or search servers. Use the Developer Tools section above, which is the largest category at 81 servers.
Do I need to be a developer to set up an MCP server?
For most servers, no. Adding one to Claude Desktop or Cursor means pasting a JSON block into a settings file, and Claude Code takes a single command.
You do need to be comfortable finding a config file and generating an API key for whatever service you are connecting. Remote servers are the easiest, since there is nothing to install locally, and only 93 of 250 servers offer one.
Can one MCP server connect to multiple tools at once?
One server wraps one system, and a client connects to many servers at once, which is the normal setup. An AI coding session might have a GitHub MCP server, a Postgres server and a browser server all connected in the same workflow. Aggregator and gateway servers also exist that put several behind one endpoint, useful when a client caps how many connections it allows.
What is the official MCP Registry?
The official MCP Registry (registry.modelcontextprotocol.io) is the canonical index of MCP servers, where publishers register under an ownership-verified reverse-DNS namespace. That verification is what makes the first-party badge in this report meaningful, and this list is built from it.
How do I install an MCP server in Claude Desktop?
Open Claude Desktop settings, edit claude_desktop_config.json, add the server under “mcpServers” with its command, args and any required environment variables, then restart Claude Desktop. Every row in the full list above provides the exact JSON, along with the equivalent for Cursor, VS Code and Claude Code.
How many MCP servers are there?
The official registry lists thousands and grows daily. This report ranks the 250 that clear the quality gate and adoption floor: a description, a working install method, and real adoption, as of 11 September 2026.
Who created MCP?
Anthropic created the Model Context Protocol and released it as an open standard in November 2024. It has since been adopted by many AI vendors, with an open-source community maintaining the specification and the official registry.
Methodology and how to cite this data
Sources: server identity, install methods and status from the official MCP Registry; stars, licence and last-commit recency from the GitHub REST and GraphQL APIs; weekly downloads from npm and PyPI; security advisories from the GitHub Advisory Database. 7 snapshots retained for trend history.
Sample: 250 servers, 3,897,997 combined weekly downloads, 11 categories, 168 first-party, 250 with install snippets across 4 AI clients.
Ranking formula: Quality Score = 35% adoption (GitHub stars and weekly package downloads, log-scaled) + 25% maintenance (update recency and release cadence) + 15% growth (30-day trend from our own snapshots) + 15% trust (first-party namespace, open-source licence, public repository) + 10% security (known unpatched advisories).
Inclusion: automatic. Every registry server with a description, a working install method and real adoption. Deleted servers excluded, deprecated servers kept and badged. No hand-picking, no paid placement.
Download attribution rule: a server’s download count is only counted when the npm or PyPI package it points at is the MCP server itself, which we test by requiring the package name to reference MCP. Registry entries pointing at a wider general-purpose library are excluded from every download total and marked library-wide in the full list, because that count measures the library rather than the server.
In this build that excluded 26 servers and 4,015,103 weekly downloads, 51% of the raw sum. The rule is deliberately mechanical, so it is reproducible and it errs toward excluding a genuine MCP package rather than inflating a total. It cannot split a single shared package across a family of servers.
Stated limitations: adoption is a proxy. MCP has no universal usage metric, so package downloads and stars stand in for it and are never presented as install counts.
Where the upstream GitHub fetch does not return data for a server, its stars, licence and release history read “not reported” rather than zero, and any trend figure that would imply a total collapse as a result is suppressed rather than published. Coverage in this build: 246 of 250 servers with a star count, 200 with a licence, 57 with attributable download data, 241 with a usable trend (0 suppressed as artefacts). Security is “not checked” for servers with no queryable package rather than reported as clean.
On the length of this page: the pages ranking for “mcp servers” run a median of nine sections and around 2,300 words, and this one runs roughly double both. That is deliberate rather than padding.
A page whose job is to let you compare 250 servers has to carry the tables, the per-category breakdowns and the method, and a reader who wants the short version has the summary block at the top and the category sections in the middle. Where a section here exists only because the format demands it, it is short.
Cite as: zplatform.ai, “Best MCP Servers: The Complete Ranked List,” data pulled 11 September 2026.
MCP servers provide the connective layer that AI systems have been missing, and the deployment story is still the weakest part of it: most of what is published runs on one laptop. Every number here traces to a row in the tables above, and where a number is missing this page says so instead of guessing. That is the whole point of publishing the method next to the ranking: the best MCP server for you depends on what you are connecting and what credentials you are willing to hand over, and you should be able to check the working rather than take my word for it.
Liked this guide? Pin ZPlatform as your Preferred Source.
Pinning us tells Google to make our hands-on AI reviews, verified lifetime deals, and founder interviews more likely to appear prominently for you in Top Stories and eligible AI Search experiences (AI Mode, AI Overviews). Set it once, no account needed on our end.
- 500+ AI tools tested with real budgets
- Verified deals - no dead affiliate links
- Editor: Alston Antony, 15+ years in SaaS & SEO
Opens Google · takes 2 seconds
Related Posts
The 30 best AI SEO tools in 2026, tested with real money: what each costs, which ones track AI visibility and AI Mode, a...
30 AI writing tools tested across free, freemium and paid tiers, every link checked, plus which AI beats ChatGPT for wri...
61 free AI image generators tested and ranked, every tool link checked September 2026, plus which are truly unlimited an...
Which AI Chrome extension should you install? 132 ranked on real Web Store install, rating and review data, with a pick ...
What Firefox AI actually does, how to enable or disable it, and every one of the 95 AI Firefox add-ons ranked on real da...
30 social media scheduling tools tested for small business: real pricing, every tool link checked September 2026, and th...