128.9 million Docker pulls sounds like a wildly popular MCP server.
In this case, it's MockServer's entire Docker image. MCP is one feature inside it. Those pulls include people who have never connected an AI agent to anything.
That distinction kept coming up while I read the READMEs behind this list. Some entries are dedicated MCP servers. Others are familiar developer tools with an MCP mode. A few are connectors or frameworks that need a closer look before you install them.
I'm the founder of mcprush.com, and this selection comes from its Top 50 MCP ranking, using the All time snapshot from October 14, 2026. I wanted to turn those download counts into something useful: what each entry lets your agent do, and the detail I'd check before connecting it.
The list runs in rank order, from 1 to 50 by all-time downloads.
One caveat before we start: these are package downloads, not active MCP users. Registry counting windows differ, CI can download the same package repeatedly, and hosted endpoints may generate no package downloads at all. Treat the ranking as a discovery list, not a benchmark.
1 Browser Use MCP
Give an agent a browser task and it can open pages, click, type and extract content. Think of workflows that cross several screens. Its roughly 370M PyPI downloads belong to the broader Python library; MCP is one way to use it.
2 MockServer MCP
Have your agent create fake API responses, inject delays or errors, and verify which requests your app sent. It also supports mocking LLM APIs. Useful for testing the unhappy path without waiting for a real service to fail. Its MCP endpoint lives at /mockserver/mcp.
3 Firebase Tools MCP
The Firebase CLI already contains an MCP server. An agent can inspect Firestore, work with Auth users, read App Hosting logs and manage Crashlytics issues. You start it with npx firebase-tools mcp. The 88.6M npm downloads count the entire CLI, including installations that never use MCP.
4 SeleniumBase MCP
A testing and scraping framework whose MCP server lets an agent drive Chrome, read pages and capture screenshots or PDFs. The catch in the ranking: its download history starts in 2016, long before MCP existed. Keep that history in mind when comparing it with newer servers.
5 Chrome DevTools MCP
For the moment when your agent says it fixed the page and you want evidence. It can inspect the browser, read console and network activity, and record performance traces. At 61.5M npm downloads, this is the highest-ranked entry here built primarily as an MCP server.
Three of the top five entries work with browsers. That makes sense to me: once an agent changes a page, being able to open it and inspect what happened is a useful next step.
6 Snyk MCP
Gives an agent access to security scans for dependencies, code, infrastructure configuration and containers. The practical use is to inspect a change with a scanner as part of the coding workflow. MCP ships inside the Snyk CLI, so its download total covers more than MCP usage.
7 Trigger.dev MCP
Inspect background-task runs, query run data, and trigger, replay or cancel jobs. That can shorten the path from a failed task to the context needed to debug it. Its --readonly mode removes actions such as deployment and replay when you only need to investigate.
8 Scaleway MCP
Turns Scaleway CLI commands into MCP tools for cloud resources such as servers, volumes and buckets. The implementation is interesting: generate the tools from an existing CLI instead of maintaining a separate set by hand. A read-only setting keeps the get and list operations.
9 EdgarTools MCP
Search SEC filings, compare company financials and inspect insider trades or institutional holdings. The agent can work from company filings instead of relying on its memory. Its README recommends starting with the agent skill, so MCP isn't automatically the best route into the library.
10 Excel MCP
Create formulas, charts, PivotTables and conditional formatting in real workbooks. It also has a formula engine that can return computed values without Excel installed. A concrete use case: turning an exported dataset into a workbook someone can open and continue editing.
11 Scrapling MCP
Fetch a page through HTTP or a browser, then return only the elements matching a CSS selector. Your agent gets the section it needs without consuming the entire page. Its filtering includes hidden prompt-injection text, though that isn't a guarantee that retrieved content is safe.
12 Office Word MCP
Build or edit Word documents with headings, tables, images and footnotes, and convert them to PDF. The result is a file you can hand over or keep editing. One limitation worth knowing: it can read existing comments, but doesn't add them.
13 PDF Tools MCP
Read selected pages, search with regular expressions, merge files or extract a smaller document. This is a focused PDF toolkit. If your task is simply finding a clause or assembling a few pages, that narrow scope may be all you need.
14 LLM Sandbox MCP
Runs model-generated code in containers with configurable resource, time and network limits. The MCP interface can also return plots. An agent can execute a calculation or script within the limits you configure for that environment.
15 Expo MCP
Lets an agent inspect an Expo app through screenshots, device and JavaScript logs, routes and UI interactions. It talks to your local development server, so start the app first. After editing code, the agent can check what changed in the running app.
16 Supabase MCP
Inspect tables, run SQL, apply migrations, deploy Edge Functions and read logs. It can follow a database-backed feature through several parts of your project. Project scoping and read-only mode are worth using deliberately. The hosted endpoint also means its npm count misses some users.
17 Azure MCP
One server exposes operations across multiple Azure services, including VMs, web apps, Key Vault and logs. An investigation can cross service boundaries without switching integrations. The scope is broad, so the credentials you give it matter. npm is only one of its distribution channels.
18 Supergateway MCP
Solves a connection problem: your client and server support different transports. It bridges local stdio and remote protocols such as Streamable HTTP. It doesn't give the agent a new business capability; it makes an existing server reachable from a client that couldn't otherwise use it.
19 Grafana MCP
Query metrics and logs, inspect dashboards, and work with alerts or incidents. My favorite detail is dashboard summaries: the assistant doesn't have to consume the entire dashboard JSON to understand it. For troubleshooting without changing monitoring configuration, it has a --disable-write option.
20 Sling CLI MCP
For moving data between databases, files and APIs. The MCP tools inspect and test connections, run SQL, copy files and validate replication configurations. Actual replication execution stays with the CLI. That boundary matters if you're expecting the agent to complete the entire pipeline through MCP.
21 MongoDB MCP
Ask an agent to inspect documents, run aggregations or read an explain plan while investigating a query. It also exposes write operations and Atlas administration. For an initial investigation, --readOnly lets you expose the data without handing over the full set of write tools.
22 Agent Device MCP
Reads app accessibility information, taps controls, fills fields and captures screenshots or logs across supported devices. One practical detail stands out: device claims prevent parallel agents from taking over each other's simulators. That matters once several coding sessions are running at the same time.
23 GitLab MCP
Read a failed CI job, file an issue or work on a merge request without copying everything into chat. This is zereight's community server, separate from GitLab's official one. It starts with a smaller toolset you can expand, and offers a read-only permission mode.
24 Next DevTools MCP
Connects an agent to running Next.js 16+ development servers so it can inspect errors, routes and logs. The agent gets live context while investigating a broken page. Upgrade workflows have moved into agent skills, so don't expect every older tutorial to match today's toolset.
25 Repomix MCP
Packs a local project or GitHub repository into a file an AI can read. Handy when you need to give a model a coherent view of unfamiliar code. Its secret scanning helps filter credentials; --sandbox restricts access to a workspace, with some functionality tradeoffs.
26 Desktop Commander MCP
Gives an assistant terminal sessions, file editing, search and process control. It can keep a development server or SSH session running while working on other steps. The README makes a useful distinction: its guardrails are not a sandbox. Isolation needs a separate boundary, such as Docker.
27 Sentry MCP
Bring the stack trace, breadcrumbs and trace spans into the debugging conversation. The agent can investigate an issue and ask Seer for a root-cause analysis. Sentry's server is primarily a hosted service; the npm package is a separate route, including for self-hosted setups.
28 Nx MCP
Helps an agent work inside an Nx workspace with documentation and running-task output. Nx Cloud adds CI failure context. A detail that caught my eye: project-graph and generator tools are hidden in the default minimal mode because skills cover those workflows.
29 XcodeBuildMCP
Build, launch and test Apple-platform apps, interact with their UI and use debugger tools. The project is now named MobileBuildMCP, while this ranking counts the older package. Check the current installation instructions before copying a command from an older article.
30 Fiori MCP
Generates SAP Fiori elements apps and modifies existing ones, including projects backed by CAP or OData. The agent can use the framework's own tooling while making changes. It also searches documentation, downloading an embedding model on first use.
31 UI5 MCP
Scaffold an app or integration card, look up APIs, validate the manifest and run the UI5 linter. It also supplies migration guidance for moving from JavaScript to TypeScript. The appeal is specific: help the agent follow UI5 conventions while it edits the project.
32 Web Forager MCP
Search the web and news, fetch pages as Markdown, or retrieve YouTube captions. The interesting detail is its packaging: it leads with agent skills for research and fact-checking, while MCP is the fallback. Read the setup guide before assuming the server is the intended starting point.
33 Code Review Graph MCP
Builds a map of functions, imports, calls and tests, then uses it to identify code affected by a change. A useful question for it is: what else depends on the function I just edited? The graph helps the assistant narrow its reading before reviewing the change.
34 Chrome Local MCP
Lets an agent work in your own Chrome with a persistent profile. It can operate tabs, read pages and run JavaScript, with local OCR on macOS. That persistent profile is the useful part, and a reason to think carefully about which browser sessions you're exposing.
35 Gmail MCP
Search mail, prepare drafts, handle attachments and organize messages with labels or filters. This is a community project, not Google's official server, and you supply your own Google Cloud OAuth client. It can also send and delete messages, so review the granted access carefully.
36 Office PowerPoint MCP
Creates slides and adds editable text, images, tables and charts using python-pptx. You can revise the resulting deck afterward. It comes from the same developer as the Word and Gmail entries, a nice example of one maintainer covering several everyday workflows.
37 ClickHouse MCP
Lists databases and tables and runs SQL against your cluster. Queries are read-only by default. I like the README's explicit warning that checks around destructive statements mostly prevent accidents: the database user's actual grants are the security boundary. Configure those before trusting a switch.
38 Google Workspace MCP
Connects Gmail, Drive, Calendar, Docs, Sheets and other Google services. This is for workflows that span apps, such as finding a file and using its contents in a document or email. It supports multi-user OAuth and a read-only mode. It's community-built, despite the broad product coverage.
39 Home Assistant MCP
Controls devices, edits automations and dashboards, and inspects automation traces. A useful task is investigating why a home automation failed to fire. This is an unofficial server with a read-only mode; the recommended installation route isn't the PyPI package counted here.
40 CAP MCP
Two focused tools for SAP's Cloud Application Programming Model: search the compiled CDS model and search the documentation. A small toolset can be enough when it exposes the exact context an agent needs. This complements the Fiori and UI5 entries above.
41 Tavily MCP
Search, extract page text, map a website or crawl it. This fits research workflows that need several sources, rather than interaction with a web app. You'll need an API key. It also has a hosted endpoint, so npm downloads capture only part of its usage.
42 Blender MCP
Lets an agent edit a scene, inspect viewport or render views, import assets and run Python inside Blender. Save your scene before experimenting. It's a third-party integration, and the package is now called mcp-for-blender, so check the current name during setup.
43 PraisonAI MCP
A framework for building agents with roles, memory and tools, then combining them into workflows. Its MCP integration is on the client side, connecting agents to external servers. It's included in the ranking, but it isn't a standalone server you add for one particular service.
44 HLIMS MCP
A connector to a remote laboratory information system. It exposes the remote server's tools and resources rather than defining its own. This is a specialist integration for that environment, not something most developers need. Its rank alone doesn't tell you whether it belongs in your setup.
45 Chroma MCP
Manage collections and documents, then run semantic or full-text searches with metadata filters. If your retrieval system returns strange results, the agent can inspect the stored records and query them directly. It supports local, self-hosted and Chroma Cloud setups.
46 Terraform MCP
Looks up providers, modules and policies, and can manage HCP Terraform or Terraform Enterprise workspaces. That combines infrastructure documentation with operations on the environment. Tools requiring explicit approval stay disabled until enabled through configuration, so check which capabilities you're actually exposing.
47 HOL Guard MCP
Reviews agent actions such as shell commands, file access and tool calls against policy. Its MCP server can inspect status and validate or propose policies, with approval required before a proposed policy takes effect. The download count covers the broader package, including its CLI and hooks.
48 Skylos MCP
Checks for dead code, security issues, secrets and quality problems. Through MCP, an agent can analyze a repository or verify a file it just changed. Check the API-key requirements: limited analysis works without one, while the other capabilities need a free key.
49 Snowflake Labs MCP
This project is deprecated and no longer maintained. It previously exposed Cortex Search, Cortex Analyst and SQL tools. It remains in the download ranking, which is precisely why a high position needs a maintenance check. Its README directs users to the official Snowflake MCP Server.
50 Argent MCP
Drives supported mobile, TV and desktop apps, records and replays flows, checks visual regressions and profiles performance. Replay the same UI journey after making a change to see what broke. Telemetry is enabled by default; the project documents a command to disable it.
What this list changed my mind about
The pattern that stood out most was MCP arriving inside tools developers already have. Firebase, Snyk, Trigger.dev, Scaleway and Repomix expose it through their existing CLIs. Distribution is already there; connecting an agent becomes another way to use the tool.
Skills are also showing up alongside servers. EdgarTools recommends a skill first. Next DevTools moved upgrade workflows into skills. Nx hides some tools because skills cover the same work. When I evaluate an integration now, I check both paths before adding another server to my configuration.
And I put less weight on lifetime totals. The deprecated Snowflake Labs entry still makes the fifty. Hosted services can look smaller than they are. A widely installed CLI can look like a widely used MCP server. Those are reasons to read the README after finding the package.
For a Next.js project, I'd start by testing Next DevTools with Chrome DevTools: inspect the running app, make a change, then check the result. For a database investigation, I'd start with one relevant connector in read-only mode. Add the next server when you can name the task it will help you finish.
Which MCP server stays enabled in your setup every week, and what do you actually use it for? I'd be especially interested in the less obvious ones.
Top comments (0)