你应该为 Claude、ChatGPT 和 Gemini 使用的顶级 MCP

欢迎收听本播客。

Welcome to the podcast.

就在昨天，我发布了一期节目，提到我和公司AI Box合作开发了一个工具——MCP，能让Claude生成图像。

Just yesterday, I released an episode where I said that I have created a tool, an MCP with my company, AI Box, that gives Claude the ability to generate images.

等Sora退出后，ChatGPT就得靠Google的Bo3来生成视频了。

It gives ChatGPT the ability to generate video once Sora dies, and they're gonna have to use Google b o three.

它还能让Gemini这类工具借助Eleven Labs生成音频。

And it gives something like Gemini the ability to generate audio with Eleven Labs.

这样一来，你日常用的主流AI工具，比如Claude、Claude Cowork、Claude Code，都可以把其他AI模型整合进去。

So you essentially can take any of the main AI tools you use day to day, and I think Claude, Claude Cowork, Claude Code is kind of one of the most popular, and you can put all of the other AI models inside of it.

最近很多人问我，第一，MCP到底是什么？它是怎么运作的？和API有什么区别？

Now I had a lot of people asking me, number one, how MCPs work, what MCPs are, what the difference between an MCP and API is.

所以我打算做一期节目，一方面分享我在AI Box里实现的一些有趣功能，以及开发过程中的收获；另一方面，也会系统梳理一下，如果你在用Claude或其他AI工具，哪些MCP最值得尝试，它们和API到底怎么工作，区别在哪，以及怎么才能最大化利用它们。

So I wanna make an episode where I'm gonna say some of the cool functionality that I've created with AI Box and some of the cool things I've learned on the journey there, but I'm also gonna break down for anyone curious what the top MCPs are that you should be using with Claude or whatever AI tools and assistance you are building with, and how these MCPs and APIs actually work, what the difference is between them, and how you can get the most out of them.

简单说两句背景——我相信你们大多数人都知道，但为了明确概念，API其实就是让一个软件和另一个软件进行通信的方式。

So as just two seconds of background, I'm sure most of you know, but just to define the terms, and explain what an API is, it's basically the way for one piece of software to talk to another piece of software.

AI、API 和 MCP 的区别在于，MCP 是一种标准化的方式，让像 Claude 这样的 AI 助手能够发现并使用数据库或服务里的任何工具、应用或文件。

And the difference between an AI an API and an MCP is that an MCP is basically a standard way for an AI assistant, you can think of something like Claude, to basically discover and use tools, any apps, any files that are inside of a database or inside of a service.

API 呢，很多时候——也不是总是——你可以把它想象成一个万能转换插头。

So an API, lot of times, not always, it's kind of like a I guess a good analogy for this is kind of a universal adapter.

对吧？

Right?

世界上每个国家的电源插座都不一样，这就像是一个个 API。

So every country in the world has a different outlet plug that's kind of like an API.

它是一种非常具体的、用来接入软件的方式。

It's kind of like a very specific way to use their software.

而 MCP 则更像一个万能转换插头，能插进任何一种插座里。

And an MCP is more like a universal adapter that will fit into any of those plugs.

因为 MCP 的工作方式，有点像一份使用说明文档。

Because the way an MCP works is it's basically, kind of like a documents page.

它本质上是为 Claude 这样的系统，全面解释你的 API 或软件有哪些功能、该怎么与之交互，而 API 则是结构固定、非常死板的。

It's basically an explanation to something like Claude of all of the ways that your API or all of the ways your software operates and how you can interact with it versus like an API was like very structured, very rigid.

举个我自己的例子，我最近做了一个工具，帮我去发布我的播客节目，但我用了三家不同的播客分发平台。

And as an example from something I've built, you know, I recently built a tool that helped me with posting some of my podcast episodes, but I have three different podcast distribution companies that I use.

我用的是 rss.com。

I use rss.com.

我用的是 acast。

I use acast.

我用的是 art nineteen。

I use art nineteen.

它们每个平台都提供了 API，方便我把播客内容发布出去。

All of them have APIs that help me with my podcast publishing.

也就是说，我只要把一集节目上传到我的平台，它就能自动推送到这些分发渠道。

So a way for me to, you know, upload an episode to my platform and have it go and publish it onto one of those distribution networks.

但问题是，这三个 API 的使用方式、接入方法全都不同。

But the problem is every single one of those APIs, like, the way they work, the way that you integrate them into software is different.

所以这些公司，还有所有用 API 的人，都会提供一份文档，告诉你该怎么构造请求、该传什么参数、会收到什么返回值。一旦你发的请求或者收到的响应有半点错误，整个系统就会崩溃，API 就彻底用不了。

So what all of those companies do and what everyone has done with APIs, they've always kind of published a an an a documents page that explains how you structure your API calls, how you have to give them code, how they'll give you code back, and if you have anything wrong in that code that you send them or you get back, it crashes and breaks and the API doesn't work.

所以这事儿真的挺让人头疼的。

So it's been pretty you know, it's frustrating.

我刚开始用Claude做工具的时候，尤其是开发这个播客分发工具，经常直接说：‘Claude，这是我的API密钥，还有他们文档的链接’——以后我会解释为什么这做法很糟糕。

And a lot of times when I was first building with, tools with with Claude and building, you know, that particularly that use case, that podcast distribution tool that I created, I would literally say, hey, Claude, here's my API, which I'll tell you why that's a bad thing in in the future.

我会说：‘这是我的API密钥，这是他们文档页面的链接。’

But I'd say, here's my API key, and here is a link to their documents page.

你得把整页、好几页的文档全读完，才能搞明白他们的API到底该怎么用、怎么集成。

Read this entire page of context, pages and pages, to figure out how the heck their API works and to integrate it.

但就算这样，也经常不成功。

And it still wouldn't work all the time.

我们得反复做大量测试。

We'd have to do tons of tests.

有时候还得联系他们的客服，一边看自己哪里出错了，一边听Claude给的建议。

Sometimes I'd to get on with their customer support, and then have to look at what I was doing and looking at my code errors, and Claude was telling me stuff.

总之，简直是一场噩梦。

And, anyways, nightmare.

所以现在有了一个很棒的解决方案，叫MCP，他们给你一个链接，你把这个链接给Claude，它就能彻底搞懂你的API是怎么运作的。

So then we have a beautiful, you know, solution, which is an MCP, which is basically they give you a link that you can give Claude that thoroughly helps it understand how your API works.

说个新鲜事，MCP本身很多时候也是在调用API，但它本质上是个通用适配器，用标准化的方式告诉AI代理该怎么正确使用这些API。

And and news flash, MCP is also leveraging an API a lot of the time, but it's basically the universal adapter to explain to the AI agent how to successfully use these APIs as a standardized way to do it.

现在你不用再费劲去琢磨API、摸索软件怎么用了，直接把MCP的链接丢给你的AI代理，它自己就能把一切都搞定。

So so now instead of struggling with an API and trying to figure out how to work with a software, you just give the MCP link to your AI agent, and it will figure it all out for you.

你再也不用去折腾那些细节了。

You don't have to finick with anything anymore.

顺便说一句，在Claude里设置MCP其实特别简单。

And it's really easy to set up an an MCP, by the way, in Claude.

我以前真没意识到这一点。

And I I didn't really realize this.

我觉得我们很多人都做过类似的事，比如在Claude或者ChatGPT里打开连接器面板，你会看到一堆提示，说‘这里有20个软件能和Claude、ChatGPT或Gemini对接’。

I think a lot of us have probably done this before with Claude or with ChatGPT where you go to the connectors panel, right, and it shows you, like, hey, look, here's like 20 different softwares that integrate with Claude or with ChatGPT or Gemini.

说实话，我第一次在Claude里看到这些连接器列表的时候，还挺失望的。

And I I'll be honest, like, I first saw all of the list of connectors on Claude, I was kinda disappointed.

我当时就想，哦，这里有二十个东西。

I'm like, okay, there's 20 things here.

我可能就用三个，比如Gmail、Slack和Google Drive，但其他那些乱七八糟的我根本用不上。

I might use like three of them, like Gmail and Slack and Google Drive, but like I don't really use all of these other random things.

而且我平时用的很多软件，其实根本没列在这儿。

And there's tons of, you know, software that I like and I use, but it wasn't listed on there.

对吧？

Right?

比如说我的邮件通讯软件Beehive，压根就没在列表里。

Like for example, my email newsletter software, Beehive, like it's just not you know, they didn't have that.

所以Beehive自己搞了个专属的MCP，你可以拿到一个链接，但它不在Claude的连接商店那种地方。

And so Beehive actually went and created their very own MCP, and it's a link you can get, but it's not inside of the Claude connections kind of, kind of like a store or whatever.

其实操作起来特别简单。

So what you can do, it's actually super easy.

你从Beehive那儿复制那个链接，点‘添加自己的连接’，粘贴进去，屏幕上就会弹出窗口，让你登录Beehive账号，你也可以用同样的方式操作aibox。

You grab that link from Beehive, you click add your own connection, you paste that link in, and it's gonna have a pop up come up on your screen that asks you to log in to, in your account on Beehive or you could do this also with aibox.

所以你把 AI Box 的 MCP 链接粘进去，就会弹出一个窗口，让你登录你的 AI Box 账号。

So you put the aibox MCP link in as a pop up that says log in to your aibox account.

登录方式就跟用 Google 账号登录差不多。

You log in just like a like a login with Google kind of thing.

这样一来，系统就完成认证了，它拿到了你的登录信息，以后就用这些信息来操作。

And all of a sudden, that is how it authenticates, and now it knows that, it has your login credentials, and it's using those.

它有了这个 MCP 链接，就知道怎么跟你的软件互动，接下来一切就自动跑起来了。

And it has this MCP link, so it understands how to interact with your software, and you're off to the races.

这时候，你什么都不用再做了。

At this point, you don't need to do anything.

不用去碰什么 API，也不用搞清楚这些技术细节。

You don't need to mess around with APIs or figure out how anything works.

你只要直接对 Claude 说：嘿，

You could just tell Claude, hey.

去 Beehive 帮我看看我那封通讯的分析数据，或者去 AI Box 帮我弄个工具，快速自动生成缩略图。

Go to beihive and, you know, help me figure out my analytics for my newsletter that I've been publishing or go to AI Box and build a tool that can help me automate, generating thumbnails much faster.

顺便提醒一下，如果你在用 Claude、Claude Cowork 或 Claude Code，而且想用上你所有那些工具的话。

Just a PSA, if you use Claude, Claude Cowork or Claude Code, and you'd like to be able to work with all of the tools that you have.

你用的每款软件，很可能都有对应的 MCP 接口。

There's probably MCPs for all of the different software you use.

你只需要手动去连接一下就行。

You just have to go and manually connect it.

你去 Google 搜一下，比如 Vercel 的 MCP 链接，找到后手动添加进去。

You just have to go Google, you know, like Vercel MCP URL, and you go get that and manually add it.

总之，希望这个小技巧对你有帮助。

So anyways, hope this was useful.

说实话，有了这个功能，我用 Claude 能做的事直接翻了十倍。

Honestly, it has 10x'd what I've been capable of doing with Claude.

拿 AI Box 来说，你只要加上它的 MCP 链接，登录认证一下，以后只要在 AI Box 里创建了任何工具——比如图片生成器、通讯稿生成器，或者音频视频生成这类工具——你都能随时让 Claude 调用它们。

And so in the case of AI Box, what you can do with it, you go and add the AI Box MCP URL, you go and authenticate and log in, and then you're able to anytime you build like a tool on aibox, if you've built like an image generator or like a newsletter generator or, you know, some sort of like audio or video generation, you know, creating ad assets like a tool, you know, you could build those kind of no code tools on aibox with all the different AI models.

如果你已经做了这类工具，就可以对 Claude 说：嘿，Claude。

If you've created something like that, you can go to Claude and say, hey, Claude.

我得去生成十张新图片，用于我的产品XYZ的Facebook广告活动。

I need to go and generate like 10 new images for my, Facebook ad campaign for my product x y z.

你去AI Box上用我的那个工具，报一下工具名称，让它生成十组新内容，记得按我们新想的方向、风格或主题来，或者结合这个新产品，它会通过MCP连接自动调用AI Box里的工具。

Go and use my, you know, my my box over on AI Box, and you say the name of the box to generate 10 new things and make sure that they have like this new direction or this new style or theme that we're thinking of or this new product, and it will go and access that box on aibox because it has the MCP connection.

它会帮你生成素材，完成后通知你，并提供链接，你可以直接下载图片、视频，或者你创建的其他内容。

It will generate your assets, and then it will let you know and then give you links back to your assets where you can go download the images or videos or whatever you've created.

如果是文字内容，我们甚至可以直接把结果插入到聊天窗口里，让你一眼就看到。

Or if it's text, think we might even port it straight into your chat message so you can see it right there.

这就是AI Box的工作方式。

So that's how it works with aibox.

我对这个功能真的特别兴奋。

I'm really super excited about it.

顺便说一句，如果你还没试过这类工具，MCP真的很有用，特别是当你需要反复迭代大量生成内容的时候。

And by the way, if you haven't tried out, something like this, MCPs are really useful specifically if you have to do, you know, generations that have many, many different iterations.

举个例子，我最近在做一个圣经研读应用，希望用户点击任何一段经文，都能让它帮我讲解背景、解释含义、提供上下文，把这段经文讲清楚。

And so an example of this, I was recently creating a Bible study app, and I wanted it so you could tap on any verse in the Bible and get it to, you know, like, give me the background, give me the explanation, give me the context of this verse, like explain this verse to me.

我经常读一些东西，完全不知道讲的是啥。

I feel like very often I'm reading something, I have no idea what it's saying.

对吧？

Right?

所以我让Claude为圣经里的每一节经文都生成了解释。

And so what I did is I had Claude go and generate an explanation for every single verse in the Bible.

问题是，旧约和新约加起来有三万到四万五千节经文呢。

Now the problem with that is there is, like, 30 to 45,000 verses in the Bible, in the old and new testament.

这工作量大得离谱，生成这么多内容耗时极长，我还希望每一条解释都得是高质量的。

It is insane, and and generating that takes a very long time, and I wanted these to be really high quality generations.

我讲到了历史背景、文化脉络，还有各种细节。

I talked about, like, the historical context and, like, all this stuff.

对吧？

Right?

于是我就搞了这么一个超大的项目。

So I had this massive project.

我想让 Claude 来搞定，但问题是，比如我说，行，去生成这些内容吧。

I wanted Claude to do it, and the problem was I would say, like, k.

它就说，好，大概得花两天时间。我就想，行吧，那就每隔几个小时看看它有没有完成一批，然后继续让它跑。

Go do these generations, and it'd be like, okay, it's gonna take me like two days to do them, and I'm like, okay, you know, I just have to keep like, continue, continue, continue every like couple hours when it does a new batch.

但问题在于，有时候电脑会关机，有时候网络会断，有时候 Claude 那边连接也会断开，服务器也会中断。

The problem was sometimes my computer would get shut off, sometimes the internet would cut out, sometimes Claude would just, you know, like, their connection would sever, their servers would sever.

有时候还会遇到速率限制，诸如此类的问题。

Sometimes I would hit a rate limit, whatever.

中途跑着跑着就崩了，原因多得数不清。

There's a 100 different reasons why halfway through a generation of a batch, it would just crash.

一旦崩了，之前生成的内容基本就全没了。

And what happens when it would crash is I would basically lose the entire generation.

可能跑了一个小时，到第50分钟突然崩溃，之前生成的一半甚至全部内容都白费了。

So it could be running for an hour, crash at minute fifty, and I would lose like half or all of everything it was generating.

真的挺让人抓狂的。

It was very frustrating.

所以我现在能做的，就是在 AI Box 里建一个盒子，内容就跟你要给 Claude 的那个 MD 文件一模一样。

So what I'm able to do is go create a box on aibox where it's it's just like the MD file that you're giving Claude to do the project.

实际上，你可以直接把那个 MD 文件拖到 AI Box 里，再用 vibe builder 加个提示词，就能生成这个工具。

And in fact, you can take that MD file and bring it over to AI Box and just give it a prompt in the vibe builder to create that tool.

我专门在 AI Box 里建了一个盒子，用来生成这些诗句描述，然后在 Claude 那边说：嘿，我需要生成三万条这样的诗句描述，就用这个盒子来处理。

But I created a box in AI Box that can generate these verse descriptions, and then over on Claude, say, hey, look, I need to go generate like 30,000 of these verse descriptions and use this particular box to do it.

这个方法有几个好处。

And a couple things it's capable of doing a couple things.

第一，因为处理是在 AI Box 上运行的，不用再等两天才生成完，它完全可以同时开启上百个、上千个，甚至两万个实例，一次性并行生成所有内容。

Number one, because it's running the processing on aibox, you can instead of having to wait two days for it to generate everything, it technically could open, you know, a 100 or a thousand or, you know, 20,000 instances of aibox and get it to generate all of them in parallel at once.

所以第一点，这能省下大量时间，不用让 Claude 一个一个慢慢处理。

So number one, that saves you a ton of time rather than having to wait for Claude to do it one by one.

你完全可以叫 AI Box 同时处理成千上万条。

You could just have AI Box, you know, do a million of them in parallel.

当然不是真的一百万条，但一次性并行处理海量任务是完全没问题的。

Not literally a million, but a ton of them in parallel.

第二点，如果其中任何一次生成失败，尤其是Claude中途崩溃了，AI Box 依然保存着你所有的成果。

And number two, if it if any of them crashes or particularly if Claude crashes in the middle of any of these generations, AI Box still has all of your assets saved over there.

比如说，Claude 崩溃了，哎呀，糟了。

So, you know, let's say, like, Claude crashes, like, oh, shoot.

我这边出问题了。

I crashed.

你可能会说，没事。

You're like, okay.

太好了。

Great.

因为 AI Box 早就把所有内容都生成好了。

Well, AI Box has them all generated.

里面有个文档文件夹。

There's a a documents folder.

AI Box 上还有一个文件管理区域。

There's a file section over on AI Box.

你实际上可以随时访问 AI Box 上所有在各个盒子或工具里生成的内容。

You can actually go and access everything that has been created on any boxes or tools on AI Box.

所以，即使我生成了大量内容，也不会有任何损失。

So nothing is lost in the case of a lot of my generations.

当你要生成一段文字时，就算失败了也无所谓，但我之前用 Eleven Labs 做过类似的事——生成了一大堆睡眠历史播客节目。

And it's not the end of the world when it's, you know, trying to generate like some piece of text, but I was doing something very similar with eleven Labs where I was generating like a whole bunch of sleep history podcast episodes.

这是我刚开始尝试的新项目。

It's a new thing I was trying.

工作室里有人想搞个新点子，我们就一起做了这些播客。

Someone in my studio wanted a new project.

我们生成了一大堆睡眠历史播客，每集大概两到四个小时。

So we generated a bunch of sleep history podcast episodes.

当时我手上有十一实验室给的大量免费额度，价值好几千美元。

They're like, you know, they're like two to four hours long, and I had a ton of, you know, free credits from 11 labs that they were giving me, and they were worth, you know, thousands of dollars.

结果我花了大约一千二百美元，自己都没意识到——因为我让我的 AI 生成工具直接去生成这些播客，然后存到指定文件夹里。

And so I I spent about $1,200, unbeknownst to me because I I told, like, my my AI generation tool to, like, go generate these podcasts, put them in a specific folder.

我当时在用Claude，生成过程中它突然崩溃了。

I was using Claude, and it crashed at some point in that generation.

我这边有个API，结果全崩了。

I had like an API, but it all crashed.

我就问，这事儿现在啥情况？

And I said, you know, what's the status of it?

他们说：哎呀糟了，生成这些内容的时候系统出问题了。

And it's like, oh shoot, yeah, we crashed generating all of these things.

你的Eleven Labs API也崩溃了。

Your eleven Labs API crashed.

所以我们现在啥都没保存下来。

And so we actually don't have anything saved.

我就问，那到底生成了多少个了？

And I'm like, okay, well, how many, like, had we generated?

我觉得它可能是卡在死循环或者出bug了。

I think it had got stuck on a loop or a glitch.

出问题了。

Something happened.

结果呢，我们在 Eleven Labs 上花了大约一千二百美元的额度，但全都崩溃了，一点都没留下来。

And it's like, okay, well, we actually spent about $1,200 of credits on 11 labs, but yeah, it all crashed and we don't have any of that.

白白损失一千二百美元，这经历真够糟心的，纯粹因为系统崩溃，而且完全没法恢复。

So that was a very painful experience to just lose $1,200 for no reason, I mean, for the crashes and it was unrecoverable.

如果你用 AI Box 来做，比如建个盒子，里面放个 MD 文件，写清楚我要做的那种睡眠历史播客的风格，你只要输入单集标题，它就能自动生成整集内容和音频文件。

If you had done that with aibox, right, so created a box on aibox that had a MD file, had an explanation of the type of, you know, the sleep history podcast that I was gonna create, and all you have to do is just give it the title of the episode, and then it will write the whole episode and generate it and give you your audio file.

要是用 AI Box，根本不会出现崩溃的情况。

If you'd done that with AI Box, then there wouldn't be the crashing.

就算 Claude 崩了，AI Box 也能稳稳生成并把文件交到你手上。

And if Claude crashed for any reason, AI Box would have safely generated and be able to get you your asset.

所以我不用再说‘糟了’。

So instead of me saying, oh, shoot.

就算 Claude 崩了，我那一千二百美元的 Eleven Labs 生成内容也照样安全存在 AI Box 里。

I just lost $1,200 of 11 Labs generations, they would all even if Claude crashes, they would all still be safe on AI Box.

所以我们现在用MCP想要打造的就是这样的系统。

So this is what we're trying to build with the MCP.

我们本质上是想让你的AI内容工作流更稳定，同时，无论Claude或者其他你用的工具出现任何崩溃，只要你的核心引擎运行在AI Box上，你的所有成果就永远不会丢失。

We're basically trying to make it your pipelines of AI content much more reliable, and at the same time, we're trying to make it so that if any any crashes happen on Claude or on any of the software that you're vibe coding, if you have the engine of it being run on AI Box, you never lose your assets.

你永远都不会失去对成果的访问权限。

You never lose access to your assets.

当然，大家特别喜欢的一点是，如果你在制作生成图片、音频或视频的工具，Claude本身根本做不到这些。

And, of course, the other thing that people love about it is if you're creating tools that generate images, audio, video, Claude just can't do that.

它不具备这样的功能。

It doesn't have the capabilities.

但如果你在AI Box上创建一个能生成这些内容的模块，你就可以让Claude调用它来完成任务。

And so if you create a a box on aibox that can generate any of those assets, you can tell Claude to, call it and to do it.

这样一来，你实际上就能用Claude生成图片了——而这是Claude以前完全做不到的。

And, essentially, you're able to generate, you know, images with Claude, which was never a capability that Claude had before.

这能为你节省大量时间，不管是为你的应用生成图片，还是为博客文章配标题。

And it saves you a ton of time for, you know, whether that's generating images for, you know, graphics for your apps or for, you know, blog post titles.

你要是去看看 aichatdaily.com 我那个AI新闻网站，所有标题、横幅图片和文章都是AI生成的。

Like, all if you go look at aichatdaily.com, my AI news site, all of the title or all of the banner images and all of the articles are AI generated.

用AI Box这类工具来生成内容，能保证风格特别统一，也不会出现崩溃丢数据的问题。

And using something like AI Box to generate them helps it to be super consistent, and then we don't have that crashing issue.

明白了。

Okay.

而且我们还让ChatGPT和Claude具备了生成图片的能力，这可是它们以前根本没有的功能。

And of course, also we we give ChatGPT or Claude the ability to generate images, which it it didn't even have before.

好的。

K.

我想聊聊目前最受欢迎的一些MCP工具。

I wanna talk about some of the top most used MCP tools that are out there.

其实挺有意思的，现在有一些MCP排行榜，能清楚看到哪些工具才是真正实用的。

There's really interesting there's some interesting leaderboards, MCP leaderboards, which are great places to see what ones are actually, very useful.

有时候公司推出一个MCP，但功能并不完整，没能覆盖他们平台的全部能力。

Sometimes a company will come out with an MCP, but it doesn't have all of the capabilities, of their platform.

说实话，我不太想挑Beehive的毛病，毕竟我特别喜欢用它发邮件，但它的MCP只提供平台数据。

And, you know, I hate to I hate to bag on, like, Beehive because I love that platform for email sending, but their MCP only gives you data about the platform.

它还不能帮你直接发帖到平台，不过他们说正在开发这个功能。

It doesn't actually help you post to the platform yet, and they're it said, like, they're working on it.

有时候，这些工具的功能还没做到跟你在网站上操作一样完整。

So there's some, you know, that sometimes the functionality isn't completely there of everything you could do if you were just on the website.

不过我想聊聊我平时最爱用的几个MCP，也说说其他一些比较热门的。

But I did wanna say some of my favorite MCPs I use and talk about some of the other more popular ones.

我用得最多的第一个就是GitHub的MCP，我觉得它好像已经集成到Claude里了。

The number one is that I'm using is the GitHub MCP, and that can also I think it's, like, built into Claude.

你很可能已经用上了这个功能。

So you probably see that one.

你大概不需要额外配置连接，但如果你要做开发，我建议你把Vercel或Netlify的MCP、Supabase、还有GitHub都连起来，搭成一套完整的体系。

You probably don't have to go get a custom connection for it, but I have everything if if you're gonna be developing anything, I recommend getting something like Vercel or Netlify MCP and getting Supabase, and then getting GitHub and connecting that stack.

这样你就能登录所有账户，想开发什么项目，Claude都能帮你直接建好数据库。

You can essentially log in to all your accounts, and anything you wanna build, Claude can go and set up your database.

它能帮你设置托管服务，还能管理代码存储位置，也就是 GitHub 和你正在处理的项目。

It can set up your hosting, and it can set up, you know, where you store your code, which is GitHub and what you're working through.

所以这三项基本上是我个人最常使用的几个主流工具。

So just those three are are basically the most popular ones that I personally am using.

我还特别喜欢那些能连接 Google 的工具，比如自动处理 Gmail、日历和云端硬盘，这些也都很受欢迎。

I love some of the Google connection ones that are able to do something like Gmail, Google Calendar, and Google Drive and interact with all of those for me, and those are also super popular.

另外，网上不少人推荐的几个实用工具里，有一个是 Brave 搜索。

And then a couple that I was seeing people recommend online that are really useful are one of them being Brave Search.

它算是目前最流行的网页搜索类工具之一，AI 能突破训练数据的时间限制，直接查到最新的开源库版本或者安全漏洞信息。

It's I think one of the most popular web search MCPs, basically the AI get can past any sort of training cutoff and can find current library releases or CVEs.

Figma 也有这样的功能，在设计和代码之间协作时特别有用。

Figma has one, which is interesting when it comes to working between design and code.

Canva 也推出了类似的功能。

Canva has some version of this.

还有 Notion，很多人在用它。

And then I think Notion, a lot of people use Notion.

如果你有这个功能，就能很好地把AI和你的工作空间连接起来，同步你的知识库、文档和各种任务。

If you have that, it's a great way to connect AI to some of your workspaces for syncing your knowledge bases, your documents, and a bunch of your tasks.

现在能接入这个功能的地方多得是。

Now there's tons that, tons of places that you can plug this into.

你可以把这些工具都集成到Cursor、Claude桌面版、Gemini，还有Grok，以及其他不少平台里。

You can plug these all into Cursor, Claude desktop, Gemini, I think Grok, and a bunch of others.

所以能接入这些MCP的平台和软件真的非常多。

So lots of different places to give these MCPs to and lots of different software.

但说实话，只要你用的软件是行业里常见的，哪怕是一些特别小众的，也很可能已经内置了MCP。

But honestly, chances are, if you have software that you use specifically in your industry and it's not, you know, some super niche software, and even honestly, a lot of super niche software even too, it will have an MCP built out for it.

所以去查一查吧，直接搜一下‘你的软件名 + MCP’，看看有没有，有的话就直接连上。

So go check, just Google MCP for your software, see if it's there, and go and attach it.

这样你就能省不少事，比如你用Claude Cowork的时候，当然你也能绕开它，但没必要。

And basically, you'll, you know, save yourself if you're using something like Claude Cowork, like, yes, you can get around this.

我经常遇到这种情况，比如它会说：‘我没有权限编辑这个项目’，其实我明明已经连了Supabase数据库的MCP，它还是提示没权限。

And I do this all the time where, like, it'll say, hey, I don't have, like, the even sometimes I have the MCP, for example, to Supabase, my database, and it's, hey, I don't have, like, the permission to edit this one particular thing.

你能帮我登录一下吗？

Can you get me log in?

或者那个特定的凭证密钥，我真不想自己动手折腾这个。

Or this, like, you know, specific credential key, and I don't really wanna have to mess around with it.

你看，我浏览器里已经登录了，直接控制我的浏览器屏幕，点几下自己搞定就行，AI也能自己完成这些操作。

I'm, look, it's just logged in on my browser, control my computer screen on my browser, go click and figure it out yourself, and it can go do that.

所以这是一种绕过某些MCP和登录限制的方法，但如果你不想每次都这么干，尤其是用了很多不同工具或者有重复性工作流程，那最好还是接入MCP，这样不仅能减少你的令牌消耗——毕竟让AI盯着屏幕、不断截图、模拟操作，会用掉大量令牌。

So that is one way to get around some of these MCPs and logins, but if you don't want to have to do that and if you're using a lot of different things or you have repetitive workflows, getting the MCPs and having it connect with them, number one, uses less tokens for you because computer use where it's like looking at your computer and moving it around and screenshotting it all the time, that's super it uses a ton of tokens.

所以如果你想节省令牌，我建议你用MCP。

So if you wanna use less tokens, I would get the MCP.

如果你想让流程更稳定，我也建议你用MCP。

If you wanna make it more consistent, I'd get the MCP.

它出错的概率更小，更少出问题。

It has less issues and glitches.

如果你在做自动化流程或者其他重复性任务，我强烈建议你为所有你用的工具都配置好MCP。

And if you're doing, you know, pipelines and other things that are repeated tasks, I would definitely get MCPs for anything you're using.

好的。

Alright.

希望这一期对你有帮助。

I hope that was a useful episode.

如果你想使用 AI Box 的 MCP，每月只需 8 美元，去 aibox.ai 就能订阅。

If you wanna get access to the AI Box MCP, it's $8 a month over on aibox.ai.

注册一个账号后，在设置里就能找到 MCP 的链接。

You sign up for an account, and then in your settings, you can grab the MCP link.

你可以把链接发给 Claude，它会帮你配置好，这样你就能访问在 AI Box 上创建的所有工具了。现在通过这个方式，Claude、ChatGPT 或 Gemini 都能接入 80 多种 AI 模型。

You can go to Claude, give that to Claude, and it will get that, get that set up so you can access any of the tools you build on AI Box, and you get 80 plus different AI models now built into Claude or ChatGPT or Gemini using that.

如果你想了解详情，可以去 aibox.ai/mcp 看看。

So if you wanna go check that out, it's aibox.ai/mcp.

我专门做了一个页面，简单介绍了它的运作方式和连接方法。

That's a page I created that will explain a little bit about how it works and how you connect it.

页面上还有一张示意图。

There's a little diagram on there.

希望这能帮到你，彻底释放出你在Claude上想实现的所有功能。

Hope that is incredibly useful, unlocks all the capabilities of what you'd like to do with Claude.

祝大家接下来的一天都顺心如意，我们下期再见。

Hope you guys all have a fantastic rest of your day, and I'll catch you all in the next episode.