I am a big fan of the Model Context Protocol. MCP keeps getting better, and I really think it's the future.
This list is my favorite MCP servers that I use every day. I picked them based on how useful they are, and how much I've been able to get out of them.
I also picked them based on these four signals:
Here is the video version of the list:
You need an MCP-capable coding agent or editor to use these servers. Kiro, Claude Code, Cursor, VS Code, and other clients support different combinations of local and remote MCP connections.
Keep in mind that whenever you install an MCP server you should be cautious. Don't accept all changes, look at what it's executing first. It's also not a bad idea to try them out on non-production code bases first, just in case.
Here are my top MCP servers, in no particular order.
| # | MCP server | Best for | Access level to watch |
|---|---|---|---|
| 1 | Context7 | Current library documentation | Low, but verify package resolution |
| 2 | Chrome DevTools MCP | Browser debugging and performance | Full connected-browser access |
| 3 | Playwright MCP | Repeatable browser automation | Authenticated pages and local services |
| 4 | GitHub MCP Server | Repositories, issues, and pull requests | Repository and workflow writes |
| 5 | Serena | Semantic code navigation and editing | Project files and language tooling |
| 6 | Supabase MCP | Database and project workflows | Schema, migration, and data changes |
| 7 | AWS MCP Server | AWS documentation and APIs | The caller's IAM permissions |
| 8 | Vercel MCP | Deployments, logs, and analytics | Project and deployment operations |
| 9 | Cloudflare API MCP | Workers, R2, D1, DNS, and APIs | Broad account API access |
| 10 | Excalidraw MCP | Editable architecture diagrams | Diagram content sent to the service |
**Project:** [upstash/context7](https://github.com/upstash/context7)
**Package:** `@upstash/context7-mcp`
Context7 is one of my favorite MCP servers. I use it all the time to help me look up documentation for my favorite APIs. It's also nice because I don't have to install 100 different skills, I can get it all in one MCP server.
When I checked, Context7 had 61,446 GitHub stars, 2,960 forks, and 3,975,539 npm downloads during August 2026. Version 4.0.4 shipped on August 28.
Why it is good: It addresses one of the most common AI coding failures without requiring broad access to your project.
Project: ChromeDevTools/chrome-devtools-mcp Chrome DevTools MCP is my browser-debugging pick. It can inspect console messages, network requests, screenshots, Lighthouse results, performance traces, heap data, and the browser state the user actually sees.
It had 50,274 GitHub stars and 3,538 forks when I checked. The npm package recorded 3,308,596 downloads during August 24 through August 30, 2026. It is an official Google and Chrome DevTools project, and version 1.8.0 shipped on August 25.
Why it is good: It lets the agent debug the running application instead of declaring success when the code merely compiles.
Project: microsoft/playwright-mcp
Package: @playwright/mcp
Playwright MCP is the browser-automation and testing choice. It had 36,660 GitHub stars, 3,077 forks, and 23,749,682 npm downloads during August 2026. That was the strongest directly attributable package-download signal in my research.
I go back and forth between the Chrome Devtool and Playwright depending on the situation.
Chrome DevTools goes deeper on live debugging, network behavior, memory, and performance. Playwright is a cleaner fit when the job is navigating a flow and proving it works repeatedly. Like check out flows.
Why it is good: It gives an agent a repeatable way to test user flows instead of relying on a one-time manual check.
Project: github/github-mcp-server GitHub's official MCP server is the obvious source-control pick. It can search repositories and code, read issues, work with pull requests, and inspect workflows.
The repository had 32,626 stars and 4,872 forks when I checked. Version 1.11.0 shipped on August 25, 2026, and GitHub supports hosted, container, and binary connection paths.
I use this all the time to push up code, or fix merge issues. I go back and forth with this MCP server and using the gh
CLI tool.
Why it is good: Source control context appears in almost every software project.
**Project:** [oraios/serena](https://github.com/oraios/serena)
**Package:** `serena-agent`
Serena is the codebase specialist. Instead of treating a repository as a pile of text files, it uses language tooling for symbol-level retrieval and semantic edits.
It had 28,686 GitHub stars, 1,934 forks, and 160,865 rolling-month PyPI downloads reported on August 31, 2026. Version 1.7.0 shipped on August 9.
I have not used Serena a lot, but from what I can tell it's really useful.
Why it is good: It gives agents a structured view of code when generic file search is not enough.
**Project:** [supabase/mcp](https://github.com/supabase/mcp)
**Package:** `@supabase/mcp-server-supabase`
Supabase MCP connects the agent to project, database, schema, and migration workflows. The repository had 2,883 stars and 403 forks, while the npm package recorded 439,749 downloads during August 2026.
This is where the list becomes very tool specific. If you are using Supabase, use this tool!
Why it is good: It gives a Supabase project direct backend context, but only projects already using Supabase need it.
**Documentation:** [AWS MCP Server](https://docs.aws.amazon.com/agent-toolkit/latest/userguide/mcp-server.html)
**Toolkit:** [aws/agent-toolkit-for-aws](https://github.com/aws/agent-toolkit-for-aws)
AWS is the cloud option I wanted to spend time on it because it's my favorite! The Agent Toolkit for AWS is larger than one MCP server. It combines MCP configuration with skills, plugins, and agent integrations.
The managed AWS MCP Server can search current AWS documentation and call AWS APIs across more than 300 services. The skill tells the agent how to approach a task. The server supplies current information and runtime tools. Same ecosystem, different jobs.
Check it out!
Why it is good: It combines current AWS knowledge with live cloud inspection while retaining familiar IAM and audit controls.
Documentation: Vercel MCP Vercel MCP is a hosted server for teams already deploying on Vercel. It can inspect projects, deployments, build and runtime logs, analytics, and traces. It can also create deployments and perform collaboration operations.
Be aware that vercel/mcp-handler is a framework for building MCP servers. The AI SDK includes an MCP client. Vercel also ships an agent plugin. None of those is the hosted Vercel MCP server.
Why it is good: It brings deployment state and production diagnosis into the agent for projects that already run on Vercel.
**Project:** [cloudflare/mcp](https://github.com/cloudflare/mcp)
**Catalog:** [Cloudflare MCP servers](https://developers.cloudflare.com/agents/model-context-protocol/mcp-servers-for-cloudflare/)
Cloudflare's API MCP covers Workers, R2, D1, DNS, deployments, and roughly 2,500 Cloudflare API endpoints.
Vercel and Cloudflare are alternatives in this list. Pick the server for the platform that runs your project rather than installing both because they appear in a roundup.
Why it is good: It exposes a broad platform API without filling the context window with thousands of tools at once.
Project: excalidraw/excalidraw-mcp Excalidraw is an amazing drawing tool, and it's worth adding to this list. Its first-party server had 5,212 stars and 487 forks when I checked. Version 0.3.2 shipped in February 2026, with code activity cited through March.
I use it all the time to help me create architecture diagrams!
Why it is good: It turns project context into an editable artifact that can leave the editor and help a team communicate.
Canva MCP is useful when you need thumbnails, launch graphics, presentations, or other design assets. Canva also has a separate Dev MCP for people building Canva apps and integrations. It does not publish a comparable server-specific adoption number, and it sits farther from the code, so I treat it as a bonus rather than one of the ten core coding picks.
Figma MCP is the design-to-code alternative. It can supply design context, Code Connect mappings, assets, and diagram information. It makes sense when the implementation starts from a Figma design.
If you need design, check both of them out! Here are a few other MCP servers I left out.
The historical @modelcontextprotocol/server-puppeteer package is deprecated, and its source lives in the
It still recorded 28,716 npm downloads during August 24 through August 30, 2026. That is evidence that package downloads can outlive maintenance, not a reason to recommend it for a new setup. Use Chrome DevTools MCP or Playwright MCP instead.
The Knowledge Graph Memory server is useful for learning how basic MCP persistence works. Its maintainers describe the reference servers as educational examples rather than production-ready products.
The package recorded 74,597 weekly downloads, but the roughly 89,000 stars belong to the entire modelcontextprotocol/servers
repository. They cannot be assigned to Memory alone. Every connected client can also read or change its local JSONL graph, which may accumulate sensitive project data.
TanStack Intent discovers and packages versioned Agent Skills. I like the project, and it is relevant to AI coding, but it is not an MCP server, client, or registry.
The distinction matters:
AWS and Vercel ship products across both categories, which is why the names can blur together. Useful does not automatically mean MCP.
I use many of these MCP servers every day. Try a few out and see how you like them.
Today my starter stack looks like this:
Add Excalidraw, Canva, or Figma when the work needs to become a diagram or design.
The full set of research links and measurement windows is available in the video description. Which MCP server has stayed in your setup after the first week?