Today, while learning about Skilljar, I discovered a very suitable MCP for front-end use—Playwright (also developed by Microsoft). I suddenly realized the significant impact of MCP on the development experience of Claude Code. In the evening, I searched and found that YouTuber Sean Kochel recommended five "must-have" MCPs, providing quite detailed introductions. However, whether they are useful or not can only be determined through testing, so I decided to write a review article.
📚REF#
Through intelligent document management, the REF MCP server can significantly reduce token usage by 85%, ensuring that the language model only loads the context most relevant to the current task, thereby improving efficiency and reducing costs.
An MCP server to stop hallucinations with token efficient search over public and private documentation.
Review#
I took a quick glance and felt that it essentially helps users perform RAG on documents to enhance the LLM's memory strength and understanding of the documents. It comes with some common documents, such as Next.js, Docker, N8N... and supports uploading private documents (Github Repo/PDF).
In fact, it is a paid project, and to use it, you need to configure an API Key. The basic functions are just two: search_documentation and read_url, and using them deducts a bit of credit. New accounts start with 200 credits, which do not refresh. What surprised me is that searching the knowledge base does not consume credits. My first experience was not great; I couldn't find any useful information. I asked for "suggestions for optimizing my project based on the Next.js documentation," but it didn't return anything and deducted 1 credit.
Next, I imported my Obsidian backup from GitHub as knowledge base material and asked, "What kind of person is the author, and what areas can be improved?" It wasn't until I pointed out that I needed to search from the private knowledge base that it finally got on the right track. The summary provided was indeed as I expected.
At this point, I suspected that my usage method was incorrect; I might need more specific questions that don't require reading the entire document. So I tried another one: "Help me find out how to maintain a block explorer from my private documents?" The results were still quite general, consuming a bunch of credits for this? 😐
Well, maybe my approach is still not right. I'll try again when I have the opportunity.
Summary#
Originally, this was a pretty good open-source MCP, and the idea is very good, but the billing seems a bit stingy, making it feel like they are using open-source as a gimmick to attract users. It would be better if users could run RAG services themselves. The experience is somewhat unsatisfactory; it searches inaccurately and consumes credits. This MCP might be more suitable for integrating some niche partners. The starting point is actually to help users save tokens, but it costs money, which puts it in a delicate balance. The MCP is only worth using when the remaining tokens exceed the cost of buying this service. However, if the monthly fee is for Claude Code, one wouldn't care so much about the tokens spent on reading documents. In the end, it might be better to manually perform RAG, find the corresponding documents, and just send a link for CC to check.
🛡️ SEMGREP#
As a powerful security scanning tool, SEMGREP has over 2000 vulnerability rules and can deeply understand code context to identify and prevent serious security vulnerabilities and coding errors that could lead to project crashes.
A MCP server for using Semgrep to scan code for security vulnerabilities.
Wow, it actually requires payment as well. At this point in my research, I began to doubt that this video was a promotion. I quickly looked at others, and aside from PIECES PLAYWRIGHT, they were all paid.
The core is still a project focused on code security auditing, but the MCP documentation is quite rough. After configuring the TOKEN, it still couldn't run, so I gave up.
🧠PIECES#
This developer memory graph tool can record and resolve recurring issues, acting as a "brain" for developers, providing instant, relevant solutions and context when facing similar challenges.
https://pieces.app/features/mcp
The code is not open-source but is free. The installation package is 1.6G, so I gave up.
🔎EXA SEARCH#
A technology search engine optimized for developers, EXA can provide the latest, most relevant technical documentation and best practices, helping developers quickly find precise information to solve complex technical problems.
Exa is a Web Search API | This is Exa MCP (Model Context Protocol)
It actually feels quite similar to REF, but it seems cheaper than REF. However, I still couldn't use it, so I gave up.
🎨PLAYWRIGHT#
An AI-powered UI testing tool, PLAYWRIGHT can automatically evaluate user interfaces based on UX/UI standards and style guidelines, providing intelligent feedback to help developers optimize the user experience and visual consistency of applications.
As expected from Microsoft, with 75k stars, it is indeed very useful and a must-have for front-end development. It will open a separate browser page for debugging, eliminating the impact of plugins.
Playwright MCP server
I'm tired; only PLAYWRIGHT, which I already knew, made me feel a bit of value. I feel like I wasted 2 hours. I think I need to reflect a bit and write a personal selection of MCPs. This article is a bit of a mixed bag, but it is what it is; it can also serve as a guide to avoid pitfalls for everyone.