My AI Stack for 2026

Every single day I get asked what AI tools I’m actually using. In this post, I’m going to show you the stack that I’m using in 2026 — the things that are actually on my computer and running my videos, my code, my programming projects, and a big chunk of my business. I’ve split this into seven categories: agent harnesses, models, editors and IDEs, my dev stack, productivity tools, coding frameworks, and AI platforms. There are over 20 tools on this list, so you’ll probably find something new here.

AI Agent Harnesses

When I say agent harness, I mean the tools that actually run AI models. The three I use are Claude Code, Codex, and Hermes Agent, and I reach for each of them for different tasks.

Claude Code

Claude Code is my daily driver. I use it multiple times a day. I have the $200/month plan, and I use it for anything I want to get done quickly since I’m running the desktop app on my computer and ask it questions, have it analyze documents, help me script out videos, and handle relatively small coding projects. Also have all my MCP servers, connectors, and a whole bank of information about me loaded in there, so it can usually answer things and knock out daily tasks fast.

Codex

Codex is the next harness I use. I don’t reach for it as often, mostly because I’ve been on Claude Code for a while and it’s my default. But I do use it for more complex development work when I don’t want to dive into all the code myself. I find it’s better at verifying that something actually works — the way it tests software feels more thorough, and it tends to give me a higher quality result. For daily productivity tasks, though, I don’t really use it, even though it performs fairly similarly to Claude Code there.

Hermes Agent

The last harness I run is Hermes Agent, spun up on a virtual private server. It handles all of my finances and accounting: tracking emails, recording invoices, receipts, and transactions, and uploading all of that to a dashboard. So not all of these are for development — most of what I run is just for running my life. I’ve tried a few other setups before landing here, and I find this one the most user-friendly across all my devices.

AI Models / LLMs

This is probably where I’ll get the most pushback, but I use a lot of different models depending on the task.

Anthropic (Claude) Models

The first family I use is from Anthropic — Claude. I mostly reach for these when I’m starting a project from scratch, want something quick, or I’m doing general business or document-type work. Even though they can be pricey, I find they perform well and get me to a first result fast. That’s not the only thing they’re good for, but it’s typically when I reach for them: something fast, quick zero-to-one, not heavy development work.

OpenAI Models

The second family is from OpenAI. These are what I reach for on complex development work — bugs, larger projects, more context to keep track of, more complicated logic. I find these models verify their own output better, and when I’m going from a working version to a more polished one, they tend to outperform the Claude family for me. That’s not a knock on Claude for these tasks, it’s just where I personally reach for each one.

Open Source Models

Next, open source models. I use these when I need long-running tasks that would otherwise burn through my subscription quickly. Right now that’s Kimi and GLM. I don’t use them often, but whenever I know a task is going to loop a lot or eat through credits, I’ll switch to one of these inside Claude Code, or bring it into a tool like Cursor. Honestly, one reason I don’t reach for them more is just that everything is already set up in my existing tools, and switching over settings doesn’t always feel worth the benefit.

Local Models

Last, local models — ones I can run entirely on my own machine. The only family I really use here is Qwen, and on my hardware I can typically run up to around 27 billion parameters. I only do this if I need to fine-tune a model for something specific, or if I’m somewhere without internet, like on a plane, and still want to get productive coding work done. I rarely use them, but when I do, I find them capable enough for a short stretch.

There are plenty of other models I’ve played around with beyond these — these are just the ones I use most. For example, I’ll typically point people toward Minimax if they’re running Hermes Agent on a tight budget and doing a lot of tool calls.

Editors and IDEs

This section is really only relevant if you’re a developer.

Cursor

My main code editor is Cursor. I find the harness inside it stronger than Claude Code or Codex, and I can switch between models or even plug in my own or local ones. I use Cursor for serious development projects — things I’m going to maintain and scale, with thousands of files. For something simpler, I’d just reach for Claude Code instead.

PyCharm

The next editor I use is PyCharm, and like Cursor, mostly for Python-heavy projects — web development, data science, or AI and machine learning work. I like PyCharm when I actually need to read through the code closely: a large repo, something I need to maintain carefully and can’t afford to break. A real IDE gives me tooling that helps me fix bugs faster than just blindly prompting a model.

CodeRabbit

I also use CodeRabbit for code reviews. I find it works well and reviews code automatically. There are other tools I use here too, but these are the main ones.

Languages and Tech Stack

Here’s what I actually reach for when building an application from scratch in 2026. Since AI is doing most of the building, my stack looks different than it would have a few years ago when I was writing everything by hand.

Node.js and TypeScript

I almost always reach for Node.js for web-based applications, which is about 90% of what I build now. It’s easy to deploy, and the codebase stays simpler when AI is building it out, since the backend and frontend are somewhat coupled. Keeping everything in TypeScript also means I don’t have to juggle two languages or explain both to the model.

Supabase

For the backend, I’ll almost always connect to Supabase. It’s easy to integrate into AI projects and comes with its own AI connector — I can view the full table schema, connect it directly to an AI tool, and deploy easily. It also handles cloud functions, triggers, indexes, role-based access controls, policies, and authentication. That means I’m not manually juggling API keys or standing up a separate database and auth service.

Supabase has been a long-term sponsor of this channel, so a shout-out to them for supporting this piece. That said, I’d have mentioned them regardless, since I genuinely use Supabase across a lot of projects. Typically I just connect to it straight from Claude Code — copy in the prompt, and it sets everything up without me needing to touch the dashboard.

Python

I also use Python fairly often, mainly for purely backend-based work or things like bots. I’ve built a number of Discord bots recently, which AI makes pretty easy, and for that I use the discord.py library.

Deployment

For deployment, depending on the project, I’ll either use a dedicated virtual private server or something like Vercel. Having Supabase, Node.js, and Vercel connected together makes deployment simple, and I can do almost all of it from my AI code editor — Cursor, Claude Code, whatever — rather than adding friction by going to a dashboard.

Productivity Tools

These are general AI tools I use that boost productivity or connect to the tools I already have.

Wispr Flow

I’ve relied on Wispr Flow for over a year now. It’s AI voice dictation software: a keyboard shortcut transcribes what you say directly into a text box wherever your cursor is. It works inside pretty much any tool, including natively in Cursor and other AI tools, so you can even tag files by voice. I like it because I speak faster than I type, and the dictation is accurate. I have a long-term partnership with them, and it’s free to try — if you use it, I don’t think you’ll want to go back.

MCP Providers

There are two MCP providers I like that let me connect all my tools into one MCP server for any AI I use: Zapier MCP and Composio. They’re pretty similar and hard to compare directly. I find Composio gives less granular control over individual tools but is easier to connect, while Zapier MCP isn’t built around discovery, meaning it adds all your tools at once, which can bloat context a bit. Both let you securely connect your tools once, then plug a single MCP server into Cursor, Claude Code, Codex, or whatever else — so switching computers doesn’t mean reconnecting 20 different pieces of software. Both also work from the terminal or CLI, and both support thousands of applications.

Obsidian

The next productivity tool I use is Obsidian. I like the wiki-style structure for storing data, and I have it connected to Hermes Agent for accounting records and vendor tracking. If you don’t already use Obsidian, it’s worth checking out.

Dev Frameworks

This section is about actually writing agents and AI-integrated software myself, rather than using a tool like Claude Code or Codex. For that, I typically use LangGraph or LangChain, especially in Python — I find they’re strong for orchestrating agents, managing context, and handling tool calling. I also lean on Pandas and Matplotlib for data science work and organizing larger datasets, and Streamlit whenever I want a simple dashboard, since it’s a lot easier to set up in Python than wiring together a separate API and frontend. Lastly, I use the official OpenAI and Anthropic APIs directly for simpler projects, since they make it easy to manage context, responses, and model switching.

AI Platforms

By AI platforms, I mean web-based platforms rather than software installed on my computer.

GenSpark

I use GenSpark for creating agents, simple automations, and workflows I don’t want to host or set up myself.

Lovable

I use Lovable mainly for generating landing pages rather than heavy logic applications, though I’ll sometimes pair it with a Supabase backend. For a quick landing page, though, it’s my go-to.

Blitz

Lastly, there’s Blitz, an enterprise AI tool I use for serious coding work. It ingests your entire codebase, and in my experience takes a few days to fully understand what you’re asking for before writing code and generating a large volume of it. Because of that turnaround, it’s really only worth it for serious projects — I find it especially valuable for refactoring or bringing something legacy up to date, since it can handle all those changes in one pass.

That’s the full stack: the harnesses that run things day to day, the models I switch between depending on the task, the editors I actually write and review code in, the languages and services I build with, the productivity layer around all of it, and the frameworks and platforms I use when I’m building something more custom. It’s a lot of tools, but each one earns its spot for a specific job rather than trying to do everything at once.