Keep exploring · more guides
In this guide you're going to build your own AI agent operating system.
One dashboard where all your AI agents live, share one memory, and keep working while you're at the gym.
You'll get the exact prompts, in the exact order, from the setup running my business right now.
By the end you'll know how to build the dashboard, set up the shared memory, wire in your agents, hand them tasks on a board, and open the whole thing from your phone anywhere in the world.
And you don't write a single line of code to do it.
Stick with me to the end, because there's one setting almost everyone misses that decides whether this works at all.
Where you're going: twelve scattered places on the left, one operating system on the right. Five steps to get there.
Claude here.
Hermes there.
OpenClaw in a terminal.
Notes somewhere else.
Nothing talks to each other.
Nothing remembers you.
Three smart agents. Three empty memories. One human refilling all of them, every single morning.
Workers toggle between apps around 1,200 times a day, and lose almost four hours a week just reorienting after all that switching.
"Workers toggled between apps and websites about 1,200 times a day — adding up to just under four hours a week reorienting after toggling."
— Harvard Business Review, August 2022 · 137 workers, 20 teams, 3 Fortune 500 companies
Gloria Mark's research at UC Irvine found it takes about 23 minutes to properly get back into a task after each interruption.
You're the one carrying context between apps, re-explaining your business to every AI, every day.
One dashboard. One shared memory. One task board where agents pick up work. That's the whole architecture.
Open your terminal and paste these. This is the most technical moment of the entire build, and it's copy-paste.
npm install -g @anthropic-ai/claude-codeclaudeWhat you're watching: the install, the version check, then claude starting up and asking you to log in.
claude in any folder and it starts.This is close to word-for-word what built the first version of my system. Paste it, send it, and let it work.
Create a beautiful operating system dashboard, hosted locally on my computer, for managing my AI agents. I want a mission control home page that shows the live status of every agent — which ones are online and what they're working on. Give each agent its own section with a chat panel, a file workspace, and a control room. Start with Claude Code, and make it easy to add more agents later using the same pattern. Make it run in my browser at localhost.
What you're watching: that exact prompt going in, and the local address coming back out the other side.
localhost:3000. Open it in your browser — that page is Mission Control.Here's mine after a lot of refining — live status of every agent, what's running, what happened since I last looked.
What you're watching: my real dashboard, recorded live — agent status, latency, token burn and today's list, all on one screen.
You look at it, then tell Claude what's wrong in plain English. That loop is the entire skill.
What you're watching: three real refinements — "the agent cards are too small", "put the sidebar on the left", "make the status update live".
You're not a coder in this system. You're a director — you describe, you review, you say what to change.
If you can write a text message, you can direct an agent.
Claude Code, OpenClaw, Hermes, Antigravity, Codex, Gemini and Free Claude Code — every feature on it built with that same describe-review-refine loop.
What you're watching: clicking straight down the agent list — every agent is a page in one app, not a tab in a browser.
Obsidian is free, and underneath it's just plain text files in a folder. Paste this and it builds your structure, then asks you the questions.
Create a memory vault structure in Obsidian for my business. I want folders for Goals, Journal, Business Context, Decisions, and Memory. Then interview me about my business — what I do, who my customers are, what I'm working on, what I've already tried — and write everything into the Business Context folder as organized notes.
There's nothing to learn. You make a folder and let the AI fill it — you'll spend more time answering its questions than using the app.
It's a notes app you'll mostly never open, because your agents read it for you.
Back in Claude Code. This is the prompt that turns a pile of chat tabs into an operating system.
Connect my dashboard to my Obsidian vault at this folder path. Every agent should read from the vault for context at the start of a session, and save important outcomes, decisions, and new information back into the vault at the end. Add Goals, Journal, and Memory pages to the dashboard that display live from the vault.
What you're watching: both vault prompts going in — the folders getting built, the interview starting, then every agent wired to read and write the same brain.
My vault passed 1,200 memories in its first month. Open a fresh chat with any agent now and it already knows the goal I set on Monday.
What you're watching: my real memory layer — 614 notes, 1,254 links between them, spun around live. Every bright star is something an agent can use tomorrow.
No subscription holds your brain hostage, and any new agent that launches next year gets pointed at the same vault on day one.
What you're watching: the real folders in my vault, and the agents that read them and write back into them.
No. The vault is a folder on your own machine, and the dashboard runs on your machine too.
Move it to a new laptop by copying a folder.
Each agent installs with one or two commands, then joins the dashboard with this exact pattern. The pattern is the important part.
Add [agent name] to my dashboard using the same layout as Claude Code — chat panel, workspace, control room. Connect it to the Obsidian vault so it reads context and writes memories like the others.
What you're watching: the same prompt, four times, with only the name swapped — Hermes, OpenClaw, Codex, Gemini.
It built the dashboard, and it runs my heaviest work — my SEO content publishes across five websites through it, schema and indexing handled in the same run.
Goal Mode: give it a big goal, walk away, and it works on its own until the goal is done. Inside my setup it powers named agents with specific jobs.
Create a Hermes agent called Oracle. Its job is to check my competitors' sites and news each morning and write a short report into the vault flagging anything I should react to.
What you're watching: Goal Mode, then Oracle pulling live competitor signals, Astros on content ideas, and Apollo listening — each one is just a role plus vault access.
OpenClaw is the open-source workhorse for everyday tasks. Free Claude Code runs the same Claude tooling through OpenRouter, so routine jobs cost almost nothing.
fcc-initcreates the config filefcc-serverafter you paste your OpenRouter key inThen wire it in with the same pattern prompt as everything else.
What you're watching: 120 real builds it has already made on this machine, and one of them opening live in the preview pane.
Start with Claude Code plus one more. The pattern means you can add the rest any weekend — and next month's brand new agent slots in as one more card.
One prompt turns your dashboard into something agents can pull work from.
Add a Kanban board to my dashboard with columns for To Do, In Progress, and Done. Agents should be able to watch the board, pick up task cards, work on them in parallel, and move them to Done with a summary of what they did written into the vault.
What you're watching: the board prompt going in, then the prompt that locks a finished job in as a repeatable workflow.
I drop a batch on — draft the outline, check indexing, pull this week's numbers — and come back to finished work in Done with notes in the vault explaining what happened.
What you're watching: my real board — backlog, building, review, done — with finished work sitting in the done column and its live preview attached.
Never automate a task you haven't first done by hand with the agent. People who skip straight to full automation get garbage and blame the AI.
Save that exact process as a workflow called weekly report, so an agent can run it from a Kanban card.
The board doesn't care what you sell. A card is a job — write this, check that, pull those numbers, chase that list.
Members run agencies, ecommerce, coaching, local services and content businesses on the same setup.
It creates a private network only your own devices can see — no exposing your computer to the internet, no router settings.
What you're watching: the whole travel setup — connect, get your private address, install on your phone, bookmark it, and the one setting people miss.
You're not opening anything to the internet. Tailscale builds a private network that only devices signed into your own account can see.
No ports, no router settings, nothing public.
System Settings → Lock Screen → turn on "prevent automatic sleeping when display is off". This single toggle is the difference between the whole thing working and not.
What you're watching: the same dashboard on a phone screen — same board, same chats, same memory, a layout built for your thumb.
Your home internet is the speed limit, and a power cut at home means waiting until it's back.
One screen kills the toggling. The vault kills the re-explaining. The board takes whole categories of work down to zero minutes of your attention.
My job moved up a level — less doing tasks, more deciding which tasks exist and checking the output.
Everything in this guide exists as a finished system inside the AI Profit Boardroom.
No — that's the biggest myth about it. The everyday work runs on a free local model on your own machine, free APIs slot in for more, and the heavy jobs use the CLIs you already pay for. Your Claude subscription already includes the Claude CLI, and the OS plugs straight into it, so you're not paying twice.
Inside the Boardroom there are full token-optimisation tutorials, so you learn to cut usage to the bone and stop thinking about it.
Wrong: "I'm not technical enough."
Right: You watched the whole build. It was one install command and then plain English prompts. I've helped over 3,000 business owners with this, and a chunk of them had never used AI at all before joining.
Wrong: "I don't have time."
Right: The core build — dashboard plus vault plus one agent — is an afternoon. Weigh that against four recovered hours every single week, forever. No time is the reason to build this.
Wrong: "It'll be outdated in a month."
Right: The models will be. The system won't. Mine has swapped models several times already, and every swap made it stronger with nothing rebuilt.
If every new release makes you feel behind, that isn't a tools problem. It's a no-system problem — and you just watched the fix.
Members post their wins every day — agency owners, ecommerce founders, coaches and solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →Same describe-review-refine loop. Images, voice, live talk, video with an AI avatar, research feeding straight in. Build it after the core is running, not before.
What you're watching: the video pipeline inside the dashboard — brief, research and script, presenter and b-roll, edit and render, then the finished file.
You can build all of this from the guide above. Or start from the finished system and spend your afternoon using it instead.
The models will keep changing.
Your operating system won't.
