👋 Hi, it’s Kyle Poyar and welcome to Growth Unhinged, my weekly newsletter exploring the hidden playbooks behind the fastest-growing startups.

Last year I invited Alex Shartsis to share how he’s playing with vibecoding to build GTM leverage. This become of the most popular newsletters of 2025. Now Alex is hooked on something new: creating his own personal AI CRM. He built it with Claude in 20 minutes for only $20 per month, and it’s already more powerful than a traditional CRM. Alex is the co-founder of Skyp, the AI-native outbound platform. He also advises companies on GTM and writes a newsletter.

Today’s edition is free. Subscribe or upgrade to support Growth Unhinged.

CRMs were one of the earliest use cases for software in business, and while they've evolved through the SaaS revolution, Web 2.0, and more modern web technologies, they are still painfully difficult to use. CRMs are mostly just databases, infrequently updated.

What’s more, typically only sales and marketing really use the CRM (if anyone does). Yet so much context exists that would benefit anyone in the organization, much of it locked up in communication tools like email.

As a bootstrapped founder and experienced CRM user/admin, I am all too familiar with the downsides of setting up a major CRM – even a modern one. But I still need to figure out who to follow up with, understand my sales cycle, and report to the rest of my team about what’s going on.

So I built out Claude to do all of this for me, and I'll tell you how to do the same thing. It’ll take you about 20 minutes (even with a hiccup or two along the way) and cost just $20 a month, all while delivering 90%+ of the value of a CRM. I'll also give you a few CRM vendors to check out that are truly AI first – in case you’d like to go that route.

The toughest challenge in this project was determining the starting point — specifically, identifying which problems I wanted a personal CRM to solve.

I settled on follow-up management because that's where I experienced the greatest context gaps and spent the most time.

I wanted to replicate the process by which I thought through follow up as a human – go back through meetings, emails, and meeting notes to figure out where things stand. Then, come up with a list of follow up tasks. Many of those would be one-to-one emails, which I would draft and send. There would also be a lot of large-scale follow up to be done – which I also would want one-to-one emails for, but by the dozens.

This led to two objectives:

I wanted to do these jobs while drawing on all of the data I already have from my emails, calendar, meeting notes, and customer list. Flexibility was important, too; sometimes I’m managing relationships around marketing (like working with Kyle on this piece) and other times I’m talking to sales leaders about software. These are very different motions, and I like keeping them separate.

Importantly, this meant I didn’t need a bunch of features you’d usually expect in a CRM (and that would’ve made it unwieldy). Things I felt were dead ends:

Briefly, why Claude? This is not an arbitrary choice. Anthropic (Claude's parent company) invented and popularized the MCP server standard. MCP stands for model context protocol.

I don't love jargon but this concept is important. MCP is a standardized protocol that lets AI models interact with external data sources and tools. Think of it as a universal adapter. Instead of building custom integrations for every service (Gmail, Stripe, Grain, etc.), MCP provides a common language that LLMs understand natively. The AI can discover what data is available and how to access it, without you writing any integration code.

This means your LLM always accesses the freshest data without you uploading anything or paying to store it twice, and without you building connectors. This is fundamentally different than "training" an LLM on your data. That is a one time event (also worth learning, but beyond the scope of this post) and there are limits to how much data you can include. Your entire Gmail archive will almost certainly exceed those limits. Instead this is giving an LLM access to fresh data, every time it needs it, without having to build extra infrastructure to do it. Find vendors that support MCP, and it just works.

To connect to your personal data sources securely, MCP currently requires a local application. The MCP servers run on your machine to broker access to Gmail, Calendar, and other services. Claude Desktop supports this today. ChatGPT does not. All of this will likely work in ChatGPT once it supports MCP, but I prefer solutions that work today, not in the future. An alternative is Cursor, which supports MCP servers, but Claude is more accessible for most people so these examples are in Claude. Anything I explain here can probably be easily implemented in Cursor in a similar way (if an MCP server connects with one LLM, it should work with any LLM).

The setup is incredibly simple, and doesn't even require vibecoding knowledge:

If you have the time, it's worth connecting your other data sources. The more you give your personal CRM access to, the more it can do for you – whether that’s simply collecting knowledge and answering questions for you, or actually making changes in your systems. Some additional systems to consider connecting:

Some of the connections are unstable, or will reset. So you may have to come back here to reconnect them. Claude will typically try to connect first, especially if you explicitly tell it “go check Grain meeting recordings” – but if it then comes back with something to the effect of "I can't access X", it won’t stop! It may just make stuff up in order to fulfil your original request. You have to stop it yourself.

While Claude is better about this than ChatGPT in that it won't invent data to finish its job, it will still do a lot of useless work that is destined to fail – costing you tokens. On your first request involving an MCP connection, always keep an eye on it and make sure it did actually connect. If it didn’t, cancel it and go troubleshoot the connection.

Many MCP servers are rate limited, just like any API. Claude usually figures this out for itself as part of the MCP protocol. What that doesn’t solve is too much context. If your dataset is very large (perhaps you have 100,000 customers in Stripe), you may exceed the “context window”. A higher-tier plan with Claude may solve this, or you can discuss with Claude how to filter the dataset to keep it within the context window.

Claude builds scripts and writes files to temporary folders that you can’t directly access as a user. If you want to reuse them, and access them directly, switch to “Claude Code” mode. This gives you visibility and access to the files and can make them more reusable.

You can also tell Claude to write the scripts to a specific folder instead of its temporary folder. This is less about cost savings and more about time efficiency. Therefore it only matters if you are using this a lot, or making the same request a lot. Even the $20 plan has enough credits where you won’t run into credit limits.

All of this is somewhat advanced. I’d start with plain Claude, as described below, and then when you see the value and have a firmer understanding of what you want done – maybe consider switching to Claude Code then.