How I use AI to build and run my projects.
Inside my setup: OpenClaw, Notion, a Second Brain and the tools that connect the work.
I am a marketer who builds things, including websites, content systems and AI workflows. The useful part is how the tools fit together: remembering decisions, finding the right context, doing the work and checking the result.
I call the operating system around this work Agent OS. My Second Brain holds the knowledge, Notion gives me a place to review it, and tools such as Codex and OpenClaw handle defined tasks. I also built my own agent, Zeus, which I control through Telegram. Here is how I have put those pieces together.
Zeus: my own agent, a Telegram message away
I built Zeus to work across Agent OS from end to end. I can give it direction through Telegram and use it to coordinate the system, carry out work and bring the results back to me. The approval and checking steps remain part of that process.
Zeus is also my day-to-day assistant for life, the universe and everything I need help with. That gives this setup a purpose beyond building websites: I want the knowledge and tools I have put together to be useful in everyday life too.
Zeus is the agent I talk to. Zeus Dispatch is the separate broker that routes work within the setup.
How the setup fits together
- DirectI talk to Zeus through Telegram to direct work across Agent OS.My own agent
- RememberObsidian holds notes, decisions and reusable instructions.My Second Brain
- Find and planqmd retrieves context. Agent OS sets the task and its boundaries.See the workflow
- Do the workCodex is the main coding tool. OpenClaw has a controlled fallback role.The agent layer
- Check and reviewTests and live checks produce evidence. Notion presents selected work for review.The review view
Decisions and useful lessons go back into the notes, ready for the next task. Notion shows a view of that work; it does not replace the underlying record.
What that looks like on Sideperk
For this site, the work started with a brief, source research and editorial rules. Codex helped turn that material into the website and its listings. The heavier tests ran on Fedora, and the finished pages were checked in a browser before release.
The first newsletter followed a separate route: prepare the listings in a reusable Listmonk template, send through Twilio Email, then check the delivered message and its links in an inbox. Keeping preparation, sending and delivery checks separate made it clear what had actually worked.
Those examples show why I look beyond the monthly price of an AI tool. The time spent setting it up, correcting its output and maintaining the connections matters too. You can borrow the approach without copying every product in my setup.
Where you can see the work
My public project portfolio covers the wider work. At MorpheOS Ltd, I develop IntelligenceOS and software for working with AI agents. These are my own projects, and I have a commercial interest in them.
The tools and their jobs
OpenClaw
Connecting an AI agent to controlled work
OpenClaw is part of my Agent OS setup for agent execution. I have documented a route in which Codex handles the task first, OpenClaw can provide a fallback for approved work, and the result comes back for validation.
What I would keep in mind
Giving an agent access to tools creates a different responsibility from asking a chatbot a question. My setup keeps approval and checking steps around its work. The diagram describes those roles, not a promise that every task completes automatically.
Notion
Reviewing priorities and completed work
I have built a Notion command centre for priorities, projects, ideas and completed agent work. It presents selected records from the underlying system in a form I can review, with links back to the supporting material.
What I would keep in mind
Notion is the review view in this setup. The durable notes live in Obsidian, and a sync needs to run before a change appears in Notion. Keeping that distinction clear helps avoid two different versions of the same decision.
OpenAI Codex
Building and checking websites
I use Codex to work on website code, investigate faults and turn a brief into changes I can inspect. Sideperk itself is one example: the work includes its pages, content handling and release checks.
What I would keep in mind
Generated code still needs testing. For a live site, I want to see the finished page and check the important actions before treating the job as complete.
DeepSeek
Building a tool for repeatable AI batches
I have built my own DeepSeek-based coding and batch tool. The workflow I have documented assigns it bounded tasks such as first drafts, outlines and structured transformations, with a separate review before publication.
What I would keep in mind
This is a builder’s workflow using the API, rather than a review of the consumer chat app. I keep spending limits and review steps around each batch; a low price per call does not tell you the cost of a finished, checked result.
Obsidian, qmd and my Second Brain
Finding the decisions behind the work
My Second Brain is the knowledge system around the work. I keep project notes, decisions and operating instructions in Obsidian, use qmd to retrieve them, and give recurring practices a consistent structure in a Knowledge Catalog. An assistant can then find the relevant context before starting the next task. This page began with those records.
What I would keep in mind
A note can go out of date. I check older records against the current project before using them as instructions or publishing a claim.
Cloudflare
Running Sideperk
Sideperk runs on Cloudflare Workers, with D1 storing its editorial records and Access protecting the editor area. It gives this small publication a place to run without a separate web server to maintain.
What I would keep in mind
Hosting is only part of the job. Domain settings, access controls and data backups still need attention. This setup is an example from my own site, rather than a cost comparison for every kind of website.
Listmonk and Twilio Email
Putting the Sideperk edit in your inbox
I use Listmonk to prepare newsletters and manage lists, with Twilio Email handling delivery. For Sideperk, that has included a reusable email template, a first edition of useful listings and a delivery check in a real inbox.
What I would keep in mind
Getting a test email delivered is the first check. Consent, unsubscribe handling and the contents of each edition need care too. I would not choose a self-hosted newsletter tool solely because the software is free.
MXroute
Handling ordinary email
MXroute handles the Sideperk contact mailbox. It keeps reader questions, corrections and replies separate from the newsletter delivery system.
What I would keep in mind
I use it for mailbox email. Sideperk newsletters go through the separate Listmonk and Twilio setup.
Fedora
Giving larger checks room to run
I use a Fedora machine for heavier builds and tests while keeping the Mac as the place I direct the work. That includes the checks for this Sideperk update after memory pressure interrupted work on the Mac.
What I would keep in mind
Moving work to another machine adds setup and transfer steps. It helps with my hardware constraint; it is not something you need just to start using AI.
Earlier use: ChatGPT for drafting and planning
My earlier work with ChatGPT included copy drafts, website ideas and spreadsheet tasks. It was useful for getting a first version onto the page and working through a problem in conversation. I still needed to check the facts and decide what to keep.
This is a historical example from my project records, rather than a claim about my current subscription or a comparison of today’s plans. My background in AI and digital work gives more context.
What this page covers
This is a selection from my documented work as of September 2026. It covers my own projects, rather than customer results or a test of every competing product. An AI-assisted review of my project notes helped assemble it, alongside my own account of building and using Zeus.
I also keep records of AI providers and models I have configured or considered. A setting alone does not show how well a product worked, so those do not get a recommendation here.
These are direct links, with no affiliate tracking added to this page. For free plans and eligibility checks, browse the offer listings. If a tool or workflow has changed, tell me.