# Connecting Your Tools

Hi everyone. Welcome to module three in which I'm gonna show you how to connect to your company tool stack or your personal tool stack. As I mentioned before, this connection is mainly done through MCPs — think about MCPs as kind of like a USB-C connector for AI. This is one connector through which AI and LLMs can interact with your external tools, such as Notion, Gmail, your CRMs, and all these tools can talk to each other. It can read data as well as write data and edit files in the destination tools.

Now I'm gonna walk you through how to connect to some interesting tools that are gonna be useful across multiple projects, and then build skills that are going to be able to use those tools in the most optimal way.

The workflow you will have is that you start your work on your local machine with the files, the markdown files, and you iterate on them. Once you're happy, you update the results in your business stack tools. So you connect Claude, and then you start sending the updates that you worked on locally to those third party tools. You can also pull information from those tools such as Jira, Atlassian, or Notion, and then again work on your machine locally and then push the updates there.

It's much easier to work on your local machine first because you can see the files, you can review them, you can do back and forth iterations, and you don't need to do all the iterations by connecting to the tools and making the changes there. You make your changes locally and then you update them in your destination tools. So for instance, if you want to craft an email, you do the draft and do the iterations on a markdown file locally, and then you say, okay, now connect to Gmail and draft this email for me there.

So now let's go ahead and connect our first tool. You can see this is the chat that I had in our project for creating the salary negotiation skill. I ran the `/context` command to show you how much context is left — 42% of it I've already used. The messages and the back and forth that I did consumed like 33% of it, for instance.

So now I would run `/clear` because what I'm gonna do here has nothing to do with what I was doing before. I clear the context and now if I run `/context`, you will see that it has much more free space — 8% is used and around 81% is already free.

So now let's go to connecting the first tool. What I want to do here is show you guys how you can give an email to your agent. This is different from connecting your own Gmail to Claude Code because we will do that towards the end of the module. Once you do that, it can create and send emails on your behalf. What I want to do here is actually create an email for my agent so it can use that email to send emails to my Gmail, for instance, or if I want to communicate with people and let them know that it's not me and it's my agent, I can use this email.

There is a company called Agent Mail, and this company provides emails for agents. To get started, you click "Get Started." I already have an account so I'm gonna sign in.

Once you do that, come to the API key section and create an API key. I'm gonna create a new one — let's say "demo project." It's optional so I don't need to fill that in. I'll create the key, copy it, and save it somewhere.

Now after having the API key, you go to the inbox and then create an inbox. Let's say "demo" as the username, so "emir-demo" is the email. "Emir Demo" is the display name. This is the email and you'll have this agentmail domain. You can also add custom domains if you want, but you need to purchase the premium account.

Now we want to see how to create the connection and give this email to Claude Code. Let's go to documentation. I'm not a developer and I'm just figuring this out with you as to how to set it up with Claude Code.

Always look for "MCP" in documentation because that's how Claude Code can connect to a tool. It already says it has MCP compatibility. I'm going to copy this page.

Many times when I implement something, this is how I do it: I go and find the right documentation. You could even just give the whole documentation to Claude and ask it to figure it out. So I go here and I'm in a fresh context. I switch to ask mode and say: "How would you implement the agent email MCP from here?"

It's asking to fetch the information. I'm gonna create another chat here and say: "Fetch web" — let me update global permissions to allow web fetch since it seems to be off by default. It should address that problem. Sometimes you need to be more explicit about the permission you want to give.

So it says it's going to do one command, install this, and get the API key from the console. We already have the API key. It's also showing a manual config and a Claude desktop config option. Let's do the Claude desktop config. It's asking if you want to do it for Claude desktop. Let me say: "Set it up for Claude desktop for me."

It's asking for the API key. I would say: "Yes, put a placeholder and I will go update it manually."

Never put your API key in the chat. It's not a good practice because it sends it to Claude through the API. So don't do that.

Now I need to check — is this the right place for having user-level MCPs for Claude? From what I know, yes. It's getting the information from there. Claude desktop is from there. There's also Claude Code CLI. This is the VS Code extension, which is what we're using.

Claude desktop is the app that I said we're not using at the beginning of the course. So I went and set it up for Claude desktop, but what I'm actually using right now is the VS Code extension. The project-level one would create a settings file in the project folder, which we already have. That's a local one. But I want to actually have it in the second option, in the user-level settings folder. The reason for this is that I want to put it somewhere that all of my projects would benefit from it, because I'm probably gonna be using agent email in many different projects. So I would say: "Do it also for Claude Code CLI."

Now this file should get updated. Let's see. It says Claude doesn't support MCP directly in the settings JSON at the user level. So my assumption was incorrect.

It says I have two options: run one command in the terminal with my API key in it, or run another command based on the documentation and it will ask for the API key. I'm gonna pick the second one. When it says "run this in terminal," you can just hit Command+J to bring the terminal up and create a new one. This is just a terminal command — Claude is not running here. So I'm gonna run this one in terminal.

It's asking to install the following package — yes, do it. So add an MCP for Claude Code. Proceed with the installation. It looks like it created the MCP locally here. Now I need to test it to see if it has access to the email. Let me reload the file and say: "The MCP is now here. Can you test it?"

Needs authentication. Open the URL, click it, and it's gonna authenticate.

It's doing it from a server. This provider is called Smithery — it's a provider that can connect across different tools. They built their MCP connection on this provider. So I'm just gonna create this account and provide the API key here. The connection should be established.

Authentication successful. Now I'm gonna say: "I authenticated it. Can you try it out?"

It has a connection. It's listing the inbox — two emails that I have. Connection is established. Let me ask: "Can this MCP live at the global level so all my projects have access to it? What is the best practice?"

So there is a global MCP JSON. It gets created when you run a specific command. The Anthropic best practice is that you can have this MCP JSON at the global level for all projects to have access if you want it everywhere. So if you want to move the agent MCP there, run this once in the terminal. This creates the MCP at the global level.

I'm gonna switch to edit mode and say: "Can you set it up at the global level for me?"

It's implementing that command. It created everything there. Then remove the MCP JSON in the local project. Let's see where the other project entry is. It's gonna remove that one too.

Let me check if it has access. I need to probably reload. Opening a new session. Let me check the context to see what MCPs are loaded — if it has the agent mail MCP. You see, it's available now. It's loaded in the setup.

Let me say: "Do you have access to the agent mail MCP?"

Now it has access and it can do all these actions. So now let's test it out.

"Can you send a test email to amir@gmail.com with a short summary of the skill that is available in this project?"

The agent is gonna send an email to me. It's probably gonna ask which email to send from because I didn't specify — there are two inboxes. It actually sent it from the demo. It listed eight available skills in the project. Now if I go to my Gmail, I should see this email. "Emir Demo" — these are the skills that are available. It came from emir-demo at the agentmail domain. "Best, your Claude Code agent."

You can change any of this — you can have your own signature, how you want to write it, all that stuff. From this, you can create a skill. I can switch to plan mode and say:

"Can you create a skill at the user level that whenever the user asks to send an email, the am-demo@agentmail.to email should be used, and the MCP that you used here should be used to send this email. And the signature should be: Best, Emir's AI agent."

The plan is ready. It basically wants to go and create this workflow. It's saying: if a user asks to send or draft an email, always use this setup. If you later on want to draft things in your Gmail, you probably don't want this skill applied, or you'd want to modify it — for instance, if you say explicitly "Gmail," it bypasses this and uses a Gmail MCP.

The skill says: always send from this email and append the signature at the bottom in the explicitly mentioned format. This is a user-level skill. I'm not sure why it went to the plugins folder, but it seems like everything on the plugins side is gonna be available at the user level for Claude Code. I'll let it do that and create it.

Now the skill got created. I'm gonna clear the context, reload the app, and check the plugins to see if it's available. The plugin is not here, but it should have access to it through the skills directory. Let me test: "Can you send me an email to amir@gmail.com? It should say hello."

It is asking for permission to use the tool. If you're fine with using the tool, you can say: "Update the permissions to allow the tool usage for this one." You can always modify permissions if you think the tool usage doesn't need your explicit approval.

I should see an email with "hello" in my inbox. Maybe it went to spam because the subject was just "hello." A spam comes back into inbox and you see: "Hello." But the email signature wasn't there. Let me check: "Why didn't you add the email signature at the bottom?"

It seems like it didn't use the skill at all because the skill was under plugins. I came back to this session and I'm gonna say: "Make sure the skill you created under plugins at the user level is available in another session. Make sure the skill is available in the correct directory."

I'm gonna move to edit mode. Let me see how it's gonna tackle it. It actually went ahead and did it despite me saying to wait, but that's okay. It removed the plugin directory, created the skill under the skills folder, removed the old email tool plugin directory, and reverted the settings. Everything is cleaned up and the skill is now created in the right place.

I can see in the skills directory we have this one. Let me go back to the other session, redo the session because the last email didn't have the signature.

I'm gonna do a clear so it doesn't have the previous context. I'll load the app again and then load the context to see if it actually sees the skill.

Excellent — on the user level it sees this "send email" skill. So now if I say: "Send an email to amir@gmail.com and include a short summary of what skills are available in this current project."

Let me go to my Gmail to check. I see the email. Eight available skills in the project. The signature is correct.

So the correct setup was that the skill needs to be in the skills folder on the user level at the root. It doesn't belong to plugins. Plugins is when you actually install something from the plugins marketplace. Anything you do manually needs to be under the skills folder. That was also a good learning for me — most of the skills I built myself so far were available on the project level, but after reading the documentation I realized there's also benefit to loading some of them across all projects, especially things like document processing.

The next tool is gonna be a tool called Reducto, which you can give any image, PDF, doc — any file that is not a markdown — and it will turn it into a markdown with high precision, even when it's handwriting or a bad quality scan. It's a document intelligence tool.

---

Now let's go set up the intelligent document processing app — it's called Reducto. You come here and this one requires a company email. Unfortunately it doesn't work with Gmail domains, so I've already gone ahead and created an account. They give you quite a large amount of credits — 15,000 free credits and then it becomes pay as you go. A single page costs about one and a half cents, so it's approximately $22 worth of credits for free and it doesn't expire.

Now you go to the API key, I already created one, so I'm just gonna use that and show you guys how to set it up in Claude Code.

This one is a fun one because it actually doesn't have an MCP. When I came to the documentation, I found a section called "Agent Guide." It says: "The complete structured reference for AI coding agents integrating the Reducto documentation processing API." I'm going to copy this page.

What I'm gonna do is go into plan mode and say:

"I want to create a skill with the scope of user, available in all of my projects, to be able to use the services that Reducto offers so I can point to any file, PDF, image, or anything that Reducto supports. The agent can connect to Reducto, send the file to them, convert it to markdown, and then bring it back. The text below is the instruction for the agent to set up this skill for me at the user level."

Then I'll add the documentation. I'll run it and pause once the plan is finished to show you the plan.

The plan is ready. It's at the right location — `.claude` at the root — so it's gonna be available everywhere. But I have a question about where I should enter my API key, so I'm just gonna ask that.

It's saying I need to add the API key in the shell environment. But I don't want to drop the API key in the chat. Let's just finish creating the skill.

Another skill is created. I can confirm the skill is in the root and it's called "reducto." It references the API documentation, which is great, and it shows what types of files are supported: CSV, Excel, presentations, images, PDFs.

For authentication, it's telling me I need to run a command in the terminal. Let me copy and paste it there and put my API key inside it.

I entered the key. "Can you test the skill to check if the authentication works?"

I got a 401 error — it's not picking up the API key. Even after restarting, Claude failed to apply it. The problem was that I could only access the environment variable if I ran it in the terminal first and then ran Claude inside that terminal session. Because I was doing it from the VS Code extension, the extension didn't have access to the terminal session, so it kept saying it couldn't find the API key.

So I have another solution. This is actually a good time to introduce a really nice app called Raycast. Let me show you quickly. This is the app — you can download it. It's a better version of Spotlight on your Mac. There's an extension called "Copy Path" which is really useful when using Claude Code. When you go to a folder and want to find its absolute path, you can just copy it from here. I'm gonna select that folder, select "Copy Path," and the path is copied.

---

Let's try a different approach. I'm gonna say: "Can you create a secret file called secret.txt? I'm gonna paste my API key for Reducto in there, and you can make sure that this skill uses that information."

It created that text file and is also configuring the skill to read it. I'm going to open and edit that file, paste my API key, and save it.

It's asking the skill to read it. It's updating the skill to load the API key from that file. Let me see if it can test now.

Now what I'm gonna do is find a sample PDF with tables. This one looks good — "Design and Build Accessible PDFs." I'm gonna download it to my Downloads folder.

Now I'm gonna point to that file and say: "Can you turn this PDF to a markdown file and store it in a new folder in the project called 'reducto'? Here's the PDF location."

You can point to a file using `@` and then mention the name if it's in your project folder, or give the absolute path. It's using the Reducto skill, connecting to it.

It's uploading. If I go to my Reducto dashboard, I can see this document came in. 17 pages parsed. Let me see how quickly it processes.

I now have a "reducto" folder with a "sample-table" markdown. This is exactly the same content as the PDF. No matter how complicated these PDFs are, this tool can process them really well and it's quite cheap.

You can use this for presentations, Excel files, anything like that — even offline files that are not easily reachable through direct connections. Upload all the relevant files, let Reducto process them, and you enrich the context of your LLM based on all these files you might have in different places. This is the best way of turning files that aren't easily readable for LLMs into markdown, which is easily parsed by LLMs.

If I do a clear and open a new session and run context, we should see the Reducto skill available. There it is. You can always say: "Use Reducto to convert this file, turn it into a markdown, and store it here." You could store it in the project or somewhere else.

Now for the next tool, I'm gonna show you how to crawl websites. This is super handy when you want to crawl newsletter websites or get structured data from a website — for gathering context when you're not able to connect directly to a tool, but you want to connect to a website and get that information.

---

Now let's go and connect FireCrawl. This is gonna help us crawl the web. There are different actions you can do: you can give it a website and find all the links on it, crawl all of them, or give it a specific URL and scrape all of it.

You can go ahead and create an account. I already have one. It can search the web, scrape the web, interact with pages, crawl sites — a lot of different things. Read about them a little bit more, but we're gonna focus on a simple use case for now.

I want to give it this website. This woman is an amazing leader who has a lot of interesting articles on how to do leadership. I just want to map all of these URLs, understand which ones are related to giving feedback, and then crawl and scrape all of them and bring them to my project as markdown files.

Now setting this up is gonna be much easier than setting up Reducto, because there is a plugin for FireCrawl already. Anything you want to connect — first look into plugins, because plugins are more off-the-shelf and you can easily connect them. FireCrawl already has a connector, so I'm gonna install it and it's gonna be available in all projects. It's gonna ask me to restart. Let me restart and run context again to see if FireCrawl is available.

FireCrawl is available here. Let me check: "Is FireCrawl plugin available to you? Can you invoke it?"

It's trying to install FireCrawl globally. It says not authenticated. It's pushing me to authenticate. I'm gonna allow it to authenticate and we should be set.

Now I'm connected. We have the remaining credits.

Now I'm gonna say: "Map all the URLs on this page."

It's invoking FireCrawl. It found all of them — all the blog article links. Now: "Can you find from her articles the ones that are related to giving feedback and just list the titles for me?"

"Can you scrape using FireCrawl the first one and store it as a markdown file for me?"

It stored the URLs in a JSON file in the project — not lost, available for future reference if I need them. I can delete the file if I don't need it.

Now it put the article here. Let me check the file. It's the one called "How to Give Better Advice" — a short one. It did crawl it.

Now I'm gonna say: "Can you scrape the next two as well?"

There we go — two more files. So just like this you can scrape many websites. There are also options for bypassing paywalls or providing credentials to access things like LinkedIn, but I'm not gonna demo that here. There are also other providers for scraping LinkedIn data using your account — one called Bright Data, and others as well. There are so many players in the market for this data scraping game that you can just go and explore them.

The best place to start is to find a provider that has an MCP, because once it has an MCP, you can easily connect to it from Claude and do the actions. You basically just need to pay for usage.

If I go to usage right now, I still have a lot of credits left and these five credits were used for the actions I just did — found all the URLs and scraped three articles. With that, now we are gonna go to implementing a YouTube transcriber API.

---

Now let's connect a YouTube transcript API. This is a YouTube transcriber which could be really useful if you want to transcribe YouTube videos. There are different things you can do with it: get a transcript, search YouTube, search inside a channel, browse channels, and also track new uploads. So if a channel sends an upload, you can take the URLs and transcribe them. You can build skills around YouTube. It's also relatively cheap to run.

Now I'm gonna show you how to set this up. You need to come here for the MCP — it says "YouTube MCP" and shows the Cloud function. I'm gonna copy this and I want this to be available at the user level. Let me do a clear here, close FireCrawl, and go to plan mode:

"I want to implement this MCP. The scope should be user so it's accessible to all projects. Also, read the documentation for Transcript API to make sure you are doing the implementation correctly."

Always resist making things complicated when you are building. It can get out of hand and you need to understand what's going on.

I just reviewed the plan. It looks decent — it creates it at the user scope in the `.claude` directory at the root. These tools are gonna be available and authentication is handled. It will automatically open a browser. Let me let it run that command.

It ran the command and didn't give the URL. Let me check with `/mcp` to see if it's listed. Maybe I need to restart. Let me reload. Checking MCPs.

I can see it here. The agent mail MCP that we created before and the Transcript AI — this is available at the user level. The status shows disconnected, so now I click on "Needs Authorization." It should open a window. Yes — I log in, I accept the authorization, and now we're successfully connected.

Let me reload and clear the context, then check the context to see if I can see the MCP. Searching for "transcript." Yes, Transcript API is there and it requires authentication — which we just did. Let me authenticate again since it's asking. It should get resolved after the first time. That shouldn't happen every time.

Let's do a quick test. I'm gonna find a short YouTube video. This one is about 10 minutes — perfect. I'm gonna take this YouTube URL and say: "Transcribe this YouTube video."

It understood it needs to use the Transcript API. It took everything and summarized it. But I want the full transcription, so: "I need the whole transcription. Store it as a markdown file in the project."

It already has the transcript behind the scenes, so it's gonna write it down. The file is created. This is the full transcript — source, author, and the full transcript. It's 16 minutes long and it got everything.

Now the next tool is probably gonna be the most exciting one for everyone, and also the most difficult one: connecting to Google Workspace. Let's see how we can figure that out together.

---

Now let's go and connect our Google Workspace account. This one is probably the trickiest to connect because it's a bit complicated and it came out quite new. I went ahead and watched a couple of videos to understand how to do it, and we're gonna do it together in the video.

In the past I used some MCP servers that had only limited access to Gmail and calendar, and some for Drive. Setting those up was also complicated, and sometimes when it was creating Google Docs it wasn't formatting them properly. But this new setup has a lot of contributors and developers behind it, including Google and Gemini — it's quite popular and a lot of people are moving to it.

Usually how I start with GitHub repositories is that I just copy the URL, paste it in Claude, and ask it to help me set this up: "Can you help me setting up the Google GWS CLI so everything is set up and I can run it in my project?"

I'm going to move to edit mode and see what happens. It's reading the repo and fetching the information.

It went ahead and checked prerequisites. Node.js was something it required, and if you don't have it, it will prompt you to install it. So I'm gonna ask it to do step one and test it once it's done.

Step one is done and tested. In the terminal I ran `gws --version` to make sure it's working. Then it has two options: Option A is the quick automated setup with gCloud, and Option B is the manual one. From the videos I watched, Option A is actually the trickier one. So let's do Option B, which involves a bit of manual work.

I'm gonna say: "For step two I'm going with Option B." It's giving me a step-by-step guide. I need to go to Google Cloud.

When you go to Google Cloud, you log into your Gmail account, then create a new project. Give it a name. I was at my project limits, so I went to one of the old projects I had. Once you create or select a project, search for "API and Services," go to "OAuth consent screen," click "Get Started," and give your app a name — let's say "Claude GWS." Put your email there.

For audience, you need to select "Internal." But this is my personal Gmail account, not a company account. If you have a Google Workspace company account, you can create an internal audience. Because I'm connecting my personal email, I don't even have that option, so I need to make it "External." With external, there might be a risk that authentication expires after a couple of days, but there is a way to refresh it. If you want to connect a Google Workspace company account, use "Internal" and add your own email.

I'm gonna add my own email to the testing mode list and save it. Now go to "Credentials," create new credentials, choose "Desktop app," name it "Claude GWS," and create. You'll have a credential file with a secret and a client ID. Download this JSON file.

Now go back to Claude. But first — one more thing: you need to enable the APIs for the services. Search for "Gmail API" and enable it. Then enable "Google Sheets API," "Google Drive API," and "Google Docs API." Once all those APIs are enabled, go back to Claude.

"I downloaded the JSON credentials. Where should I store it on my machine so it's available everywhere?"

It's giving me a location to store it. That's a safe place and all projects will have access to it. Using Raycast, I'm gonna copy the path to that file and say:

"Here's the file. Please move it to the directory you recommended."

It got removed from Downloads and created over there.

"What's the next step? Run it for me."

It gave me a URL. I'm gonna paste it in the browser and authenticate. I'm going to my account, granting access to all these things, closing the browser. Access granted. Everything in the GWS authentication is complete.

Now I'm gonna test it. I have an email that I send myself every day. I have an automated workflow that processes my second brain in Obsidian — it schedules things in my calendar, creates tasks in my to-do list, syncs my calendar with Obsidian, reads transcribed meeting notes, and a bunch of other things. Every day at 8 AM it runs and sends me an email with a report of what was processed, what it did, and if it faced any errors. I built this through time as a skill with full instructions, and scheduled it with Claude Code.

So now I'm going to ask Claude Code to return all of the emails from this sender and store them in a markdown file.

"Use GWS to connect to my Gmail and bring all the emails I received from this sender. Store them in a single markdown file in the root of this project."

Found 21 emails. It's pulling all of them. Behind the scenes it's actually writing Python code to do this — probably the more efficient way of doing it. It pasted all the emails sorted by date and is now creating the markdown file.

There we go — that's the file. These are all the emails that this agent sent me.

Now let me enable Bash permissions globally so it doesn't keep asking. "Enable Bash command usage in my permissions globally. Use the config." It should update. If I refresh to a new session, it would apply. That's a good practice.

Now let's do the next test. I'm gonna ask it to create a Google Doc in my Drive. "Connect to my Google Drive with GWS and create a Google Doc with the list of all linear tasks from this file." I'll mention the emails-from-agent file.

I hope the authentication works because if it doesn't, it'll ask you for a URL again. You click it, authenticate, and sometimes these tokens expire, especially in testing mode — I think it's around 10 days or less. But right now it should be fresh.

It couldn't figure out on its own that it had access to the GWS CLI. Let me say: "You should have access to GWS, Google Workspace CLI."

It got access. Maybe you actually need to create a skill so it knows what you mean when you say "GWS." Now it's making calls through the Google CLI, connecting. The title of the doc is going to be "Linear Tasks" and the doc is created.

Let me filter by date modified. "Linear Tasks GWS" — there's our doc.

Now: "Can you categorize them based on title?" Interesting — there's no title in the emails. So I'm gonna say: "Create a heading underneath the file called 'Meetings' and add all of the meetings there as well."

Let me see if it manages to create the heading. GWS gives you options for formatting the doc properly.

The formatting is a little bit off. Actually no — it did put the heading in. I'm gonna say: "Can you improve the formatting of this doc for better readability?"

Now the document has a title, subtitles, and improved formatting for all items. I can even ask it to create tables.

"Okay, now take the linear tasks with all their attributes and create a Google Sheet from them."

It's amazing — I didn't specify any formatting but it figured out I cared about it. It went ahead and created proper formatting for the sheet.

"Can you create a filter for status?"

It seems GWS does have access to that capability in Google Sheets. Filter added.

"Can you filter all the items that are 'Synced'?"

Wow, it killed it. So you literally don't have to touch your sheets anymore. You can do whatever you want from here. Everything that a user can do — including complicated formulas — you can build right here.

The power of this is that it has all the context. From here you can create everything. For instance, if you want to transform data from one tool to a Google Sheet, you pull that info by connecting to the MCP of your tool, get all the information to your local file, and then push that information to a Google Doc. You can build skills around it to make sure that every time you push, everything stays in sync. From there, you could even create a sheet and then from that sheet construct email drafts for you, which you can queue up and send — or even schedule emails and keep the status in sync between your Gmail and Google Sheets. Basically the limit is the sky. Anything that you can do as a person on Google Sheets and Google Suite, Claude can do for you as well.

Now the last tool, which is Notion. The interesting part here is that I'm gonna connect it to a tool that has a lot of information about me and ask a question about that database.

---

Now let's connect my Notion. I want to pull information from this database that I have, which is all of the articles that I read and podcasts that I listen to. I have an app called Readwise, which is a great app if you're a fan of reading articles or watching YouTube videos — it's like a bookmarking tool that pulls every highlight from these different places into Notion or other tools. You could actually pull it to Obsidian as well.

I want to connect this Notion and then ask: "What are the parenting advices you can find from my past bookmarks?"

Notion should be an easy one because there's a plugin for it. I'll check the plugins, install Notion, and it's gonna be installed everywhere. It'll ask me to restart. After restarting, I'll open a new session and check if it's connected.

Running context — the skills are here so it's most likely already loaded. "Are you connected to the Notion plugin? Are you authenticated?"

It's doing a quick search. Actually, it was already connected because of my previous connections from the claude.ai Notion integration. Plugins are generally very easy to connect — it would've just asked you with a browser URL, you go in and authenticate.

Let me see if it can actually see the Readwise database. "Can you see the Readwise database?"

Interesting — it's in the wrong workspace. "How can I authenticate to the correct workspace? How can I switch to it?"

"I don't want to connect through the Claude AI integration, I want to connect through the plugin using local connections to Notion."

I went ahead and switched to another account and disconnected Notion from the Claude AI integration. Now I'm gonna say: "Can you connect to Notion?"

It's saying to run the MCP. The connection is there but not authenticated — it's going to ask me to connect. You click that, and it should open a browser. If it doesn't open a browser automatically, I can ask: "Can you give me the authentication URL?"

Something to watch out for: the agent might try to use FireCrawl to crawl Notion's documentation pages. You may want to make a rule that it should not access FireCrawl without your explicit consent. Or in the plugins menu, you can actually turn FireCrawl off like that and disable it when you don't need it, so it won't use your credits.

So I know how to do it now. I'm gonna say: "Run this in terminal with user scope."

It went and created the MCP. I wanted it at the global level, and it's done. When I go to `/mcp`, I need to restart first, otherwise it won't show. After restarting and running `/mcp`:

Notion is there. "Authenticate." I come to a browser, click yes, I recognize this callback, go ahead. Now I'm fully authenticated.

Going back: "You're authenticated. Can you see the Readwise database?"

There it is — it found it. "Can you go through all the records and bring me a maximum of 20 parenting advice pieces, and put them in a markdown file in the root of this project."

It's using the MCP, making calls to the Notion tool, and trying different search approaches. It's doing: "All records tagged with parenting." Let's see if it can figure it out by itself. It's making different calls and getting information.

It found a lot of information — parenting, parents, parents. I'll see. I have a lot of information on parenting.

The document is already created: "parenting-advice." You see all the information it brought in — different sources, 20 different pieces of advice. It went ahead and summarized some of them, which I didn't explicitly want, but you can ask it to include the full content.

"How to Raise Kids Who Are Not Assholes" — let me search for this in Readwise. Yeah, the original content was actually small, so it might have summarized some of them. You could explicitly say: "I want the full content here."

You can see the context got filled quite quickly because of all this text — behind the scenes it kept the content in the context for the session. If I inspect the context, you see that 33% is messages. I didn't write a lot of messages, but all that content went into the messages, so that's why it's a bit bloated.

Now I definitely need to run clear and then do something else or compact it.

One thing to note: if you want to capture the knowledge base of your company, do not do it from the MCP side only. Use the MCP if you want to do a search or read specific files. But if you want to do a mass download, it's gonna be really heavy on your tokens — there's no need to pay for that. What you should do is use exports. In Notion settings, there should be a section to export all documents. In "General," there's "Export content." You can export as Markdown, CSV, or HTML — HTML is actually better because it includes images. Choose "Everything" and "All content." This exports your whole workspace.

Once you do that, put all those files in your project folder and now you have the whole context of your Notion available to your LLM. From there you can construct a clean context for, say, all your product requirements into one document, or a comprehensive competitive analysis. Whenever you want to do something, think about the context as something to distill to exactly what you need, without all the fluff and unnecessary information that isn't needed for the LLM to get the job done.

With this I conclude module three. The next module is the capstone project, where we're gonna bring all of these tools together in one comprehensive project: planning a trip to Japan. I'm gonna show you how I think about a project when I start it from scratch. I gave you a glimpse of it through setting up configurations, but we're gonna do a practical one together. Hope you guys enjoy it. Thank you for watching. See you in the next module.
