F2F #91: My AI stack

I've spent the last couple of years integrating AI into how I work at MarsBased, and even though things change rapidly, I feel like my AI stack is pretty consolidated... for now.

I have the feeling that somewhere around the corner lurks The Next Big Thing™, which is going to change my AI stack drastically. Maybe it's Fable, maybe it's not (we'll know when they stop playing hide and seek with it). Up until December 2025 everything revolved around ChatGPT for me, whereas now, ChatGPT has residual use in my day to day.

The core philosophy: augmentation, not automation nor replacement

When we started repositioning MarsBased around "AI-Augmented Development," it wasn't marketing smoke. It was me finally articulating something I'd been living with for a while: I don't want AI to replace us, I want AI to augment us. To give us superpowers. Take the tedious tasks so that I can focus on stuff that really moves the needle, if I so want it.

I have also seen plenty of times catastrophic approaches to marketing automation, pre-AI and now with AI, and I have yet to see a use case where it doesn't come across as cheap to the person on the receiving end.

For now, I'm not chasing the shiniest tool or the one with the best benchmark results. I'm looking for things that extend my capacities & skills without making me dumber in the process.

The tools I actually use

Claude: This is where most of the work happen. I moved off ChatGPT at the beginning of the year and rarely ever I go back to it. Claude chat for the reasoning, debating, brainstorming, document processing and pretty much everything; Claude Code for code (duh) and large documents (bigger context windows help to manage huge documents), and Claude Cowork, I barely use it because it doesn't share context with the other two, so I feel like I'm wasting my time reinventing the wheel. Further, if you use more than one device (in my case, one laptop and one mac mini), the lack of shared context layer is especially painful so I feel like I sometimes avoid working on certain tasks because I have to switch to the other computer.

Perplexity: It's been almost two years since I replaced Google with it and I don't look back. I used to go to Google for the images, but Perplexity has gotten significantly better over time in this regard. Now, I can't remember when was the last time I visited google.com to search for something. Perplexity gets all the quick queries I don't want to store. Volatile shit belongs to Perplexity whereas more meaningful prompts I want to keep around belong to Claude.

Linear + MCP integration: I live in Linear for project management and issue tracking, for work related matters. The MCP (Model Context Protocol) server for Linear means Claude can create tickets directly, read context, and understand our workflow. That bridge between my thinking environment and my execution system matters a lot, especially useful when I can dictate stuff into Claude and it turns it into Linear commands. For instance, when I need to create tasks in bulk, like "create a task for every current client we have to remind me to reach out for the quarterly follow-up calls".

Todoist: Similar to Linear, but for personal matters. This way, I can keep personal and work separated. I also manage Todoist via Claude nowadays, and because Claude has access to both of them, when I get the daily brief, it knows this logical separation so it doesn't mix up personal tasks with work-related tasks in my list of to-dos for the day.

n8n for workflows: Most of my infrastructure lives here. I have written a POSSE syndication (write once, publish everywhere) tool, reminders, notification routing & more. n8n isn't AI, but it's the glue that makes AI tools talk to the rest of my stack and I host it on Hetzner, which means it costs almost nothing and I own the data. Still learning the ropes.

Hetzner: This is the unconventional part. I run FastAPI apps for things like reminders and memory, PostgreSQL databases, and increasingly, personal tools built on Claude API. Everything is self-hosted because I care about owning my data and the cost structure of personal tools should be under ten euros a month, not subscription tiers.

ChatGPT: Surprise! I still use it! It's mainly because I paid for a yearly subscription upfront so I still have access to it, but credit where it's due: its Voice Mode is unbeatable at the moment. I can't use Claude Voice because it's downright terrible, whereas chatgpt's is not only more stable but it also has got more capabilities and feels more human. When I want to talk for 30-60 min with an AI, I still talk to ChatGPT.

Holo: I use Holo to track my health, diet, supplementation and fitness routines. For the full disclosure, I am an investor in the company.

Close but no cigar

I don't use AI for something just because I can. I've turned down interesting integrations because they don't solve a real problem I have and I also don't use more tools because I simply don't have the time for them. For instance, I've tried Grok and Gemini but they simply didn't cut it for me.

I have replaced Chrome with Comet, but I don't use its agentic capabilities because the web is so unsafe that I don't feel right by using them unless it's on the MarsBased website and a few others. It shows promise but it's too dangerous for now. Also, to be entirely honest, I tend to forget I have it because F2F #67: Similarity kills innovation so I still think I'm using Chrome instead.

I have tried a few AI-powered apps like Plantora and Gyroscope, the first to take care of my plants, the second to take care of myself, but they ended up feeling like more slop. Lazy wrappers of AI with very clunky interfaces and little care for the product itself.

I also don't pretend this stack is final because it's not. Right now, I'm thinking about how to build better memory systems, how to integrate voice more naturally, how to make sure the tools I use actually match how my brain works and how to build a unified context layer for all my tools. It's a fucking lot of work, but slowly and steady, I'll get a bit closer to my goal every day.

A few considerations

Integrations, finally! The real value isn't in any single tool, it's in how they talk to each other. MCP servers, APIs, webhooks... I can finally connect everything, cross data and get rid of the eternal blocker I always had of not being able to connect sources to get the full picture. Now I have it, so I don't have excuses other than not having the time to access the tools or the data. I can write something in Linear, Claude sees it, creates a task in my personal infrastructure, which triggers an n8n workflow, which sends me a whatsapp notification. It's not perfect but it's taken me way further than I ever reached before AI.

Control over convenience: I use Claude API heavily for my own services because for certain things, I want to have 100% of the control, especially when I feel like there's an app that's too overpriced or when no app in the market does what I exactly need. I wrote about it here in Fuck you code.

Reasonable friction: I could automate away most of my decision-making, content generation, publishing and what not, but I simply don't want it. As I mentioned above, the marketing stuff always looks cheap and cold to me, but for other things like strategic decisions or important emails, I want to think through them. If a tool makes it too easy to skip thinking I stop using it because I stop valuing the effort required to reach a great outcome, eventually devaluating the output itself. AI should reduce the tedious parts, so you can focus on the thinking parts.