Transcript
00:00:00Imagine you're a 19th century pirate and during your looting you've stumbled upon a new AI tool.
00:00:05Previously you have been using MCP tools with your AI agents. When you have an MCP tool connected
00:00:10your model has to call that tool and it gets some results back. The pirate has the Gmail MCP
00:00:15connected because he needs to email his therapist about his work-life balance. With the MCP you can
00:00:20ask Claude to find emails from a specific person and reply to them. It calls the search tool to
00:00:24find them, then gets the threads to read each one, then the reply tool to send a response. But what
00:00:29if the Gmail MCP was literally a folder and Claude could go inside it the same way it goes inside any
00:00:34folder on your computer? Your emails would be actual files that Claude reads the same way it reads markdown
00:00:39files in your code base. That's exactly what Mirage is. It's basically a virtual file system for AI agents.
00:00:45But before we dive deeper into this virtual file system let's take a look at why file systems are
00:00:50so important for these AI tools. The significance of file systems comes down to how AI models actually
00:00:56learn things. When you give Claude a custom tool, whether it's an MCP or an API, you're asking it to learn
00:01:02something new on the spot like a baby being forced to start walking at gunpoint. The agent has to read
00:01:06the description and plan which calls to make in what order. That's tokens of overhead spent before it does
00:01:12any actual work and it repeats for every tool you add. A file system doesn't have that problem because
00:01:17it's the one interface every LLM has already learned thoroughly during repetitive training. Basic file commands have been
00:01:23used billions of times in the code the model trained on. Unix has been refining this for 50 years and it's
00:01:29why file systems are the one abstraction that lets an AI agent work across multiple services without
00:01:35repeatedly forgetting its life's purpose. During thorough testing, I found lots of sneaky little things that aren't
00:01:40written inside the tool's GitHub description. After the setup, we asked Claude to mount Gmail to the Mirage
00:01:46workspace. Like a good little boy, it had to guide me through the Google Cloud platform for the OAuth
00:01:51creation so that we could get the credentials. When you start Claude inside the Mirage folder,
00:01:55you'll be able to see the services you mount. You can see that Gmail has been added here and for each
00:02:00Gmail label and category, it could see a different directory. Now before we dive into the new features,
00:02:05we gotta put in the word from our sponsor. Before you point the finger of judgment, remember the
00:02:09editor's kids need to eat too. Now if you're paying for ChatGPT, Claude, Gemini and Midjourney
00:02:14separately, you're already spending way more than you need to. ChatLLM from Abacus AI brings over a hundred
00:02:20AI models into one platform including ChatGPT 5.5, Claude Opus 4.7, Gemini 3.1, Grok and DeepSeek and
00:02:28every new model gets added instantly. You don't even have to choose because RootLLM automatically picks
00:02:33the best one for your prompt. But it goes beyond just chat. You can use it for writing and deep
00:02:38research, generating images and videos from top models like Seedance 2.0 and Nano Banana. They also
00:02:44have the Abacus AI agent. By chatting with the AI, you can build professional presentations, create full
00:02:49stack apps and websites that even accept payments so you can launch a business without writing a single
00:02:54line of code. So chat, deep research, images, videos, apps, websites, AI agents, presentations,
00:03:00applications, literally everything in one platform. All of this for just $10 a month instead of paying
00:03:05for separate AI subscriptions. Check it out at chatllm.abacus.ai or click the link in the description
00:03:11and start building now. Now there's a problem with the existing Gmail MCP. Emails regularly get file
00:03:17attachments with them but using the MCP, Claude can't read or download them which pretty much makes it a
00:03:23blind old man in this situation. It can only see the file name which creates a serious limitation with
00:03:28the automation. But with Mirage, if we ask it to search for any attachment, what's going to happen
00:03:32is that it can actually read the file attached to the email because it's just a file in the file
00:03:37system. With the MCP, you have to get the attachment manually and paste it into the agent's context
00:03:42manually. In our inbox, it found this extremely legal invoice and it read the whole thing and gave us its
00:03:47actual content. Unlike your ex, Mirage actually cares about you and gives you some predefined tools
00:03:53that you can easily add such as Notion, the whole Google suite, Telegram, Slack and even different
00:03:58storage systems or databases as well. If you connect them, they will appear as folders to Claude in its
00:04:03directory. And the same thing can happen with Slack where chats become files and any file attachments that
00:04:08your team sends there can also be accessed easily. Since Google Drive is a storage platform, its MCP will
00:04:14obviously give you a download tool for the files. But there's a huge flaw in the download mechanics. The
00:04:19d*** thing is only efficient for small things. And no, we aren't making that joke here because my
00:04:24grandmother watches these videos. For example, we had this small 2MB file that we wanted the Google
00:04:29Drive MCP to download and place somewhere else. To download the file, it fetches the file's content
00:04:35as a string and then it is decoded and written back into your file system. Even this request took
00:04:40about 4 minutes and also bloated the context. And if the file size is huge, let's say 100MB,
00:04:46then it would hit the response limit and waste the context as well. Mirage solves this because you can
00:04:50just copy the files from the drive onto your system. For example, we had this sample test data file,
00:04:56it could actually read it and it used the copy bash command to just copy it into my documents folder.
00:05:01The setup process for this is actually pretty long so we're not going to go through each step and we
00:05:06don't need to because we're not living in the olden times anymore where you actually had to read
00:05:10documentation like a common peasant. What we recommend is just cloning the source code directly
00:05:15onto the system and once you've cloned the source code, just open Claude inside that and let it guide
00:05:20you through the installation. Going through the installation steps one by one doesn't really
00:05:24matter because it's all code and Claude can automatically do everything. But if you run this
00:05:28on a Mac, the install isn't that straightforward because god forbid Apple lets you install anything
00:05:34without feeling like you're about to burn your system to the ground. To use Mirage,
00:05:38macOS needs to have it mounted as a file system so that your agent can use it as just another folder
00:05:43with files. This is why it uses this library called MacFuse. Using this software, macOS starts
00:05:48supporting third-party file systems. Claude code can automatically download it via homebrew and all you
00:05:53have to do is follow its steps. Basically, MacFuse is a system extension and you do need to restart
00:05:59your Mac and change the security configurations. So that is something that you should be aware of.
00:06:03But for those of you who are living under a rock and are new to AI agents, this might cause
00:06:08a problem for you. When it's going to restart, that chat session is going to be lost as well. So either
00:06:12rename that chat session using the built-in slash command or ask it to just read the previous chats
00:06:17because all of them are stored on the system. This is why when we restarted the system, we had to give
00:06:22it the chat again and it recalled the context that the chat was about setting up Mirage for the video.
00:06:27Most of the tools that you will use need to be authenticated. Unlike the connectors for Claude,
00:06:32automatic OAuth isn't built in here. So whatever tool you need to install, first it's going to mount
00:06:37a TypeScript file where it's going to add the tool as a directory in this workspace. After that,
00:06:42you're required to add in the credentials that the tool needs. For Google applications, you will have
00:06:46to go into the Google Cloud Console, which has the user experience of filing taxes in a foreign language.
00:06:52You have to enable APIs and get credentials and that might take you some time. It is a little time
00:06:57consuming, but Claude is going to guide you through it pretty easily. After all of this is done, Claude is
00:07:02going to ask you to run a mount command in a separate terminal and keep that terminal open. Now, if you
00:07:07don't want to keep the terminal open, you can ask Claude to run it as a background process. And after that,
00:07:11you basically just go into the Mirage folder, start Claude there. And as you can see, when we list the files,
00:07:17it has all my connectors and it can actually look inside them and search for files. Now, if you've been
00:07:22following the channel for some time, first of all, thank you. You're our favorite human being. And you
00:07:26probably remember the code mode video on Docker, where Docker's code mode could basically make custom
00:07:31tools by combining tools from different MCPs so that repeated operations could be run easily.
00:07:36This is the same thing. You can do the same just by using bash commands combined together so that the
00:07:41contents of one file can be copied into another. For example, here, what we asked it to do was find
00:07:46emails in the inbox that mentioned sponsors, get us the top three sponsors. And yes, top three obviously
00:07:51means the sponsors that give us the most dough. Then it would list them on the Notion page. And
00:07:55another thing that code mode or TypeScript tools in general solved was that context never really had
00:08:00to enter the model's context window. For example, here, the model just has to write bash commands that
00:08:05fetch one thing from one file and then put it into another. The model doesn't need to have all of
00:08:10this loaded context placed inside its context window. So the same thing happens here with bash commands.
00:08:15They form a pipeline and they can easily connect these different tools together. Now, these are not the only
00:08:21tools that you can use with this virtual file system. What's basically happening on the back end is that
00:08:26it's using APIs from these tools, the same APIs that are used by the MCP servers, and it's just turning
00:08:32those into file systems. This is like convincing the homeless man outside our building that he is the
00:08:36chosen one and he can actually fly. So the same can be done with any custom service that you want. If you
00:08:41want a Figma MCP server where all the Figma designs are listed as files, you can do that as well.
00:08:46Although we don't think people honestly use Figma that much anymore. If we recall correctly, it was
00:08:51supposedly just killed by Claude Design weeks ago. But other than that, we tried it out on the Google
00:08:56Chat APIs. Internally in our team, we use Google Chat. And yes, we want you to judge us for that.
00:09:01Google Chat allows you to go into spaces and add different apps there. So we basically made this
00:09:06test space and inside it, we added this Mirage app. It basically gave it access to read everything in
00:09:11there. And as you can see, it automatically fetched everything and was able to read all the context.
00:09:16And you can do this easily with any service. Again, as we mentioned in the setup, make sure that you have
00:09:21the repo installed locally and that Claude can read everything in there so that it doesn't have any
00:09:25gaps in its context. Of course, it's going to take some retries because just like us, Claude isn't
00:09:30perfect. You're going to get a lot of errors and you might have to do multiple feedback loops with
00:09:35the agent, but you can pretty much build any tool into a file system. Now, we already told you that
00:09:40you don't have to keep it running in a terminal. It can run as a background process, but every time that
00:09:45process restarts, you're starting from scratch. The cache and the index are what make the workspace
00:09:50persistent and both of them are gone on every restart. This is exactly what every new AI founder
00:09:55has been doing for the past two years. Mirage fixes its own problem with a demon that turns it into a
00:10:00persistent background server. You can have multiple workspaces running at the same time. They can be
00:10:05named and each one starts in whatever mode you want. But let's say you don't want this on your local
00:10:09machine or you work across multiple machines because your fat homeschooled brother can't
00:10:14get enough of Roblox on the PC. But because it's a standard HTTP server, you can host it anywhere
00:10:19and switch to your M3 Max MacBook. That's what lets you turn Mirage into a file system knowledge base.
00:10:25That brings us to the end of this video. If you'd like to support the channel and help us keep making
00:10:29videos like this, you can do so by using the super thanks button below. As always, thank you for
00:10:34watching and I'll see you in the next one.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video