The AI stack I actually use

Why the fifty-tool list doesn't help you

Ask for the best AI tools for a small operation and every result reads the same: a countdown from fifty to one, a screenshot of each pricing page, and a paragraph of praise that could have been written by whoever built the tool. Nobody in that list says which three of the fifty they would actually miss, or what would make them delete the other forty seven.

So this is not that list. It is what I actually run, sorted by the job it does, not by what a marketing page called it. Some of it I touch daily. Some of it I am still deciding about, and I say so plainly. One thing I dropped, and I say why. No affiliate links, because I do not have any to give you.

The only question worth asking before you add a tool

Every tool in a stack that is actually working has a cut condition: a specific, checkable thing that, if it happened, would get the tool dropped. Not "if it got worse," which is not a condition anyone can check, it is just a feeling you will keep postponing. A real cut condition names something you would notice, this week, if it stopped being true.

Ask most people running an AI stack why they still pay for something, and the honest answer is closer to "I set it up once and now I am afraid to touch it" than to any actual reason. That fear is not a cut condition. It is inertia with a login.

So the table further down in this guide is not a nice extra tacked on at the end. It is the actual claim of the piece. If a row in it cannot carry a real, specific answer in the last column, the tool does not belong on the list, however good it sounds in a sentence describing it.

Wiring and memory: n8n and Hermes

n8n is the layer that moves data between things and wires up the smaller agents. It is a workflow automation platform that lets you build visually and drop into real code, JavaScript or Python, when the visual layer runs out of road, with several hundred integrations already built. That combination, no-code for the common case, real code when you need it, is why it sits underneath most of the rest of this list rather than being a name on it by itself.

Hermes is where I put persistent, cross-channel agent memory. It is Nous Research's open source agent, self hosted, and it connects to Telegram, Discord, Slack, WhatsApp, Signal, email, and a CLI, carrying context across whichever one you are using that day. One of the channels I route through it is WhatsApp, where a small internal CRM I built for the real estate team I work with posts issue updates instead of sitting in a tab nobody keeps open. That is one thing Hermes does, not a tour of what the CRM is, so that is as much detail as it needs here.

Other people run the same kind of job through OpenClaw instead. Same category of tool, different choice. I picked Hermes because it was the one I got running first and because the cross-channel memory matched what I needed, not because of any argument against the alternative.

The content pipeline: Notion, autopilot research, and transcripts

Notion is where content operations live: what has been recorded, what is drafted, and what is published, tracked in one connected workspace rather than a spreadsheet nobody opens twice.

Research runs on autopilot underneath that. Claude's routines run a saved prompt on a recurring cloud schedule, hourly, daily, or weekly, without needing my laptop open, and I pair that with GPT's own scheduled tasks running a similar recurring prompt on its side. Both are moving toward Hermes as well, since it already holds the memory the rest of the pipeline needs.

For raw material, I pull from the Agent-Reach repos. It is an open project that gives an agent one-click access to platforms it would otherwise need separate setup for, including turning podcast audio into a full transcript and pulling YouTube subtitles without extra credentials. That is the front half of the podcast pipeline: get the words out, with the source attached. Once a transcript lands, it feeds the same six-row extraction table I wrote up in turn one idea into a week of content. The stack's job stops at getting the words onto that table with a timestamp attached.

Lead capture: one plain table

Airtable is where leads from my websites land today. It works like a relational database wearing a spreadsheet's clothes, and the point of keeping it there is boring on purpose: one place, already running, already connected to the handful of other tools it needs to talk to.

The contact form on this site writes to Airtable rather than introducing a new service, for the same reason. Reuse what is already running before you add a dependency. What happens to a lead after it lands is a separate job, and I cover the research and outreach side of that in find leads with AI.

Scraping without building a scraper: scrapegraph and scrapling

Scrapegraph is a scraping library that uses an LLM plus its own graph logic to turn a page into structured data. You describe what you want out of it, and it handles the extraction instead of you writing a selector for every field on every page you need.

Scrapling does the harder version of the same job: an adaptive web scraping framework built to keep working when a site's layout changes or something is actively trying to block a normal request. I run both alongside the Agent-Reach repos from the content pipeline above, since gathering raw material for research and for transcripts is the same underlying job: get the source text out cleanly, with something to blame when it comes out wrong.

Building: the ladder before the code

Ponytail is a plugin I run in Claude Code, including on this site, that forces a decision ladder before any new code gets written: does this need to exist at all, is something like it already in the codebase, does the standard library already cover it, is there a native platform feature, is there an already installed dependency, and only after all of that, is a small amount of new code actually justified. It is the same discipline as the cut condition above, just applied one step earlier, before the code exists instead of after.

Superpowers installs a skills framework into Claude Code on top of that: composable workflows for test-driven development, systematic debugging with a checkpoint after repeated failed fixes, and a review step before anything gets called finished. Task Master AI turns a longer brief into tracked tasks and subtasks so a big build does not quietly drift while I am not watching it. Context Optimization and Planning with Files round out the same group, keeping a long build's context and its plan in a form I can reread instead of trusting my own memory of a decision made three days ago.

That same discipline, one job, a defined input, a defined output, a place a human checks the work, is what I write about at more length in build an agent that does real work. For anything with a screen, the Frontend Design skill has Claude commit to a palette, a type pairing, and a specific aesthetic direction before writing a line of code, and the result ships as a Claude artifact, a live page I can hand someone a link to. Context7 sits underneath the coding itself, resolving a library name to its current, version-specific documentation so the code matches the API that library actually has this month.

Research: Tavily, Deep Research, and NotebookLM

Tavily is the search layer built for agents rather than for a browser tab: a single API for searching, extracting, and crawling live web content in a shape an agent can actually use, instead of a page full of ads and cookie banners.

Deep Research is not one specific product on this list so much as a mode I reach for wherever it is available that week. NotebookLM's version of it, for instance, browses up to hundreds of pages on its own and returns a multi-page report, which is the shape I want when I am trying to understand a market rather than answer one question.

NotebookLM itself is where documents go to get interrogated rather than just stored: PDFs, slides, sheets, web pages, even a YouTube video with captions, turned into something I can ask questions of and get an answer with a source pointing back at the actual page. I use it for property research and market reading more than anything else on this list.

Documents and brand: the boring, necessary layer

Anthropic's own document skills, for PDF, DOCX, and XLSX, cover contracts, listings, proposals, and deal sheets, the paperwork a property business actually runs on. None of it is exciting. All of it has to work every time, which is a different bar than most of the tools higher up this list have to clear.

Brand Guidelines and Canvas skills do the same job for my own content: keeping colors, type, and layout consistent across what I put out, without me re-explaining the brand every time I ask for something new.

What I'm exploring, not using

Honcho and Mnemosyne are both in the testing folder, not the working stack. I want to say that plainly, because it would be easy to let this section quietly turn into a recommendation.

Honcho is agent memory infrastructure built by Plastic Labs. Its own documentation frames memory as a reasoning problem rather than a retrieval one: it tries to draw conclusions about a person or a conversation over time, not just fetch back what was literally said. Mnemosyne OS is smaller in scope. By its own description it sits alongside Obsidian and reads the vault without writing to it, adding meaning-based search over notes that already exist.

My own reason for looking at both is that I want a memory layer that lives under whatever assistant I am using, not another app to keep open and synced. Neither has earned a place in the working stack yet. I have not decided, and I am not going to pretend I have.

What I dropped, and why

Obsidian is out. Not because it did anything wrong. My notes stopped being the actual constraint on my week, and what I wanted instead was memory that lived underneath the assistant I was already talking to, not a separate vault I had to remember to open and keep in sync. That is the whole reason. No drama, no feature I was missing, just a different shape of problem than the one Obsidian solves.

The AI tools for solo founders, laid out by what would get them cut

This is the boring, inspectable version of everything above. If a row cannot survive being written down this plainly, it should not be on the list.

What I'd actually tell another solo founder

Do not copy this list. Half of it exists because of a business you do not run and a team you do not have. The part worth taking is the test underneath it: for each tool you are about to add, can you say today what would make you drop it. If you cannot answer that in one sentence, you have not adopted the tool, you have just started paying for it.

So here is the small thing to do with this. Pick one tool you added in the last month without asking that question. Try to write its cut condition in one sentence, the way the table above does. If you cannot, you already know which of two problems you have: either the tool is not doing enough to earn a real answer, or you never actually decided to use it in the first place.