Hi Reader,

-> One quick thing - if you want to learn about this stuff in a webinar format, I'm running weekly webinars on AI, marketing, and automation topics for the foreseeable future - registration details at the bottom of the email.

I’ve had fun doing the business development dance over the last week, and filling my pipeline with opportunities big and small.

I’ve been able to play with different messages and say my “pitter-patter” out loud and narrowing my focus towards B2B companies selling services $10-30k+ that need help with marketing and automation.

But onto the main reason you read this newsletter - stories!

This week I look at:

• Exploring MCPs and deploying them with Claude Desktop
• What does your LinkedIn network say about you
• Wednesday Webinars

There’s oodles to get on with in this edition so let’s make like a backyard pool and splash on with it.

What in the MCP? Setting up and playing with a new protocol

MCP stands for model context protocol, and is a new set of rules for helping LLMs talk to other software.

Think of it like a language.

If Japanese is your native dialect, for other people to interact with you fully - they need to speak Japanese as well.

But unlike the language analogy, MCPs improve over time as more and more people use them, and the data going in and out of an MCP increases.

It’s a standard, and as the standard increases so has interest.

I’ve avoided the hype up until recently, but with Claude’s capabilities getting better and better, it was time to see what the fuss is about.

I installed a number of MCPs (official and unofficial) to see what I could do.

So here’s 3 I tested:

1. Apify
2. Airtable
3. n8n

Sidebar: I tested these with Claude Desktop and instructions on how to do this are here. There are other (and much more secure) ways to set up connections so do your own research or wait until the tech becomes more accessible.

Apify is a scraping tool that has a number of “actors” that will run basic through to complex scraping requests either as one-off runs or on an ongoing schedule.

Ever wondered how bigtime influencers choose what content to create?

Well one of the ways they do it is by scraping the web and popular sites to see what’s popular and use that information to inform what they do.

​

And here’s a snippet of the kinds of data you can pull (example from Instagram Reels):

​

This is all very grey market at best, and very clearly violating terms and conditions of these sites.

The flip of this is they’re all selling your data and feeding it into Large Language Models so I’ll leave it up to you to judge where things sit.

Apify gets technical fast.

To plug it into other tools you need to understand APIs, how to make different requests, plug things through JSON and you can get bogged down.

So when I heard Apify has an MCP I was curious about how easy it would be to pull request and track usage (it’s a paid service).

One thing that’s been bugging the sh*t out of me is the effort I will go to with a YouTube video and not have many options to repurpose it without a lot of time and effort.

So I checked how hard it would be to transcribe all 90 videos.

So I pulled my YouTube URLs from Advanced Analytics, and pasted them into Claude and asked it to.

Bang.

It did 86 videos for $0.17.

Including the transcript from this old video:

Airtable (to date) didn’t have an official MCP server - so I grabbed access to one from Github.

My CRM is in Airtable, and I wanted to know if I could update things in my CRM by just asking Claude to do it.

I can, but the problem is MCPs eat up a LOT of context window.

This is the amount of characters/tokens I can use in a single conversation window.

I found that if I asked 2-3 follow-up questions it would fill out and hit its limit quickly.

It was great for individual queries like this:

There are better ways to do this and the tech is quite new, but it’s pretty exciting.

The last tool I played with and checked out was n8n.

I had a pretty robust set up with producing n8n workflows already (custom project instructions, a lot of project knowledge, and lots of example workflows).

And I was curious just how much better the solution could be.

It turns out a lot.

When a workflow was a complex request, Claude would often produce JSON files with gaps that wouldn’t upload well.

I’m building a solution for content repurposing (creating social media snippets and posts from my YouTube videos and newsletters) and the difference was pretty dramatic.

My biggest issue was running out of context window for amends and changes to workflow requests.

My workaround for this was swithcing to Claude Code (another topic I’ll cover in the future).

So far pretty good right?

Well - here’s the bad news.

MCPs are a flat-out high-security risk when run from your desktop machine.

A developer friend sent me this tech influencer showcasing a way to exploit these tools with prompt injection.

I didn’t watch the whole thing but it’s wild seeing someone stream in a developer environment with a RØDE microphone and hollering like it’s a Twitch stream (which is already weird enough).

Our timeline is so strange and explore this part of AI at your own risk.

What does your LinkedIn network say about you?

I keep up with a lot of news thanks to Readwise/Reader.

When an article or post catches my fancy, I save it to Reader via the extension or share mechanic on iOS.

(Reader is the next iteration of Readwise - think of it like a smart read-it-later tool).

I don’t really save a lot of prompts - as usually I prefer working out what I need with iteration.

An exception caught my eye the other day:

Or just uploading the data and asking “what questions should I be asking about this data?”

I happened to do a LinkedIn data pull the other day, so this is what I did.

It gave me all sorts of information - some useful, some blowing smoke up my arse.

I then asked it to put a short hitlist into the “Prospect” stage of my Airtable CRM - and it did!

Marshy’s Webinar Wednesday

There’s a few things I love doing:

1. Presenting
2. Shooting hoops
3. Dancing to techno

I can’t get paid for the last 2 (despite much practice) and so I’m going to run hard at webinars on AI, marketing, and automation for the next couple of months.

There’s a lot of information out there and by distilling and presenting, distilling and presenting, I see a way I can learn faster, cultivate more interest in this space, earn trust, and practice what I preach as a startup consultant.

It’s also a lot easier to remix and repurpose my content and IP than it was 18 months ago - so let’s f*ckin’ go!

The details:

• A webinar on AI, marketing, and growth every Wednesday lunchtime for 30 mins (12pm AEST)
• Limited to 30 spots
• Recorded except for live Q&A
• Live Q&A available to only attendees
• First topic: 3 marketing things you didn’t know you could do with AI
• Bookable here - lukemarshall.net/webinar-wednesday​

My goal is simply to ship for the first few weeks as I figure out mechanics and run with this - request future topics here.

–

Alright that is plenty.

I’m back with a vengeance baby.

-Marshy

p.s. I’m currently working on automating a few solutions across content creation, webinars, and content repurposing - if you have a problem you think can be solved by AI but aren’t sure on the how, reply to this email and throw it down man