Transcription
All right, so this is a neat new feature that was added to the Cloud desktop app. It is the Model Context Protocol, I believe, MCP for short. Just want to run through a quick example of this.
So on the left-hand side here, I've got some files here, one of which is a SQLite database here, the test DB. If I were to open that database up with the command line here, inside of here I have a table. I select everything from that products table, got about 20 records in here. There's a price column. One thing I could do to test this out would be to sum up the prices across the item. So, select the sum of the price from products. So there's the sum.
And the reason I computed that is because now I want to ask Claude to do it. And to get Claude to do it, I need to make a small configuration to point it at that database. So right now, you can see there's no tools available in the Cloud desktop app. Got that running here on Mac OS. You can do it on Windows as well.
What I need to do is I need to come into the settings here under the developer tab. I need to edit the configuration. It's going to open up a file that I can modify here. You can see there are no servers right now. So I'm going to paste this into that file. I dump out the contents of that file again here. You can see it's now configured with a server. Now I called this SQL light. I could have named this whatever I want. You can have multiple databases in this case, and there are other tools as well. Nonetheless, I've got this configured.
So now if I come back over to Claude here, I need to restart the Cloud desktop app. So now if you look over in Claude in the lower right corner, there are six tools available, MCP tools. Click on that to see what they are. Of course, you could imagine I could query data, I could add data, create tables, etc.
So let's do this here. First off, let's just ask Claude to tell me about the database. So, "Tell me about my database." It's going to prompt me for permission. I wish they would allow me to just approve this indefinitely for this particular integration, if you will, but I haven't seen that ability yet. So I'll go and allow it for the chat. Then I'm prompted again here to describe a table. So the first time is asking to list tables. Now it's asking to describe tables. Those are two different tools it can use. Now it's asking to be able to run a query. So I'll go ahead and allow that. You can see it's selecting all from products. It's just getting a sampling of the records, of course. It's telling me about all of this here.
So now that I've got my response here, I used three different tools. You can see them here. Here's one of the tools, here's another tool. And of course, it reports back to me that there is a products table, just one table, and then here are the columns or the schema. And then last up here, I asked it to read the records. It took a sampling of those, it got back about five of them here. So it's basically describing it's a table with products with their names and prices, and it's asking, "Would you like to know anything specific? For example, the price range?" So we could ask that.
Let's do this though. Let's come up with our own question here. So, "What would it cost to buy everything?" So, "Sum up the price" like we did over on the left here. It's going to prompt me again here. Actually, no, it's not going to prompt again because I already approved running queries. It did that. All right, there you go. You can see 1642 here, and that matches over here on the left with 16428 as well.
So that's just one of the tools that's available. If I open up a browser here, you go to pypi.org and you search here for MCP, you can find other tools. Here is the protocol SDK. You can see it was just released here. And if you put in MCP and server here, you'll find more matches. For example, here's the SQLite server using. So when I configured that tool to point at the database, if you remember here, I specified this MCP server SQL light. So this is the Python package that's being used. It's got some parameters then.
So as you can see here, there are other tools packaged up. Like here are some Git tools, fetch tools, which I assume help you browse the web. I'm sure other people will be adding in packages. And I should say at this point in time, be careful with what you use in here. Anybody can add a package to pypi.org. So just careful if you're going to use one of these, what you give it access to. It's not just about what Claude can access, it's also if you're running software from somebody else to do that, what they can then access.
And so I assume there will be some page somewhere, maybe it already exists here. Let's look for the Model Context Protocol. This is from Anthropic. All right, yeah, they've got good docs here. Got a blog post here. And then I believe they have, okay, this is the documentation I remember seeing. So currently, I have not yet found a list of packages that are deemed safe to use. Use just keep in mind there should be a list at some point. You should know where the package is coming from before you use it.
And yeah, I'll definitely make more videos about this. My assumption here is this is kind of a response to what OpenAI just released, where they have the Works with Apps feature on Mac OS, where you can have the ChatGPT app connect to iTerm, Xcode, all sorts of different environments and be able to work with that environment. Assuming this is kind of the same response here, maybe a bit different, but it's going to give them the ability to connect to other applications, interact with them on your local computer. And I think there's a lot of power here, especially if you can kind of speed up the process and work on the trust factor, so it doesn't take so long to prompt for the ability to do something. If you can do all of that, I think there's a lot of automations that can be set up by opening it up to the community to be able to create their own automation packages or servers. That's going to be really incredibly powerful.
And one more thing here, I want to show you. Come over to clients here. These are tools that can use this protocol to connect to a server, for example, to do something. You can see right now we have the desktop app. We also have the ability, it looks like within Zed, to be able to do something with prompts, which is interesting. Yeah, as slash commands here. I'll have to definitely take a look at that. I'll make another video if that's something that I like. And then there's this SourceCraft Cody. I've never used it, but it's got some sort of integration. I assume this list will grow over time. And definitely keep an eye on this. I think there's a lot of value possible when it comes to letting these AI systems interact with your actual computer, your resources, even remote resources could do some really neat things for you.