CONGRATULATIONS. YOU HAVE BEEN CLEARED FOR ACCESS.
Your personal AI lives right at home, running on the Mac Studio in our living room. This guide is your survival briefing. Read it or don't — but don't come crying to Cathy when your agent calls you by the wrong name.
Sign in with the Google account that matches your name. Wrong account = the door doesn't open. That's not a bug, that's a feature.
Click any section to expand. Yes, you have to read them. The Black Knight is watching.
The Claw Tower is a private, family-run AI platform. Think of it as having a team of extremely knowledgeable assistants who live in our house, never leave, and never tell anyone what you asked them.
Unlike ChatGPT or Google, your conversations don't leave the building (unless the task is genuinely complex — that's when it calls Claude, but you'll see that happening and it's budget-capped).
The Tower is a virtual building. Each family member has their own floor. Your floor has rooms — each room is a specialised AI agent (researcher, calendar, email, etc.). Your stuff stays on your floor. Other floors can't see it.
To get in: go to tower.kiriakosai.com, sign in with Google, and you're home. No app to install. No account to create. Your Google account IS your key.
An AI "soul" is the system prompt — the foundational instructions that shape how your agent thinks, speaks, and approaches problems. It's the difference between "generic robot assistant" and "an AI that actually sounds like it understands your life."
Without a soul, your agent is a blank-slate language model. With one, it becomes your research partner, your planner, your creative collaborator — tuned to how you think and what you care about.
Navigate to tower.kiriakosai.com. Find your floor in the building view and click a room (agent) you want to personalise.
In the agent panel, click the Settings tab. You'll see a "System Prompt" field — this is where the soul lives.
Describe who your agent is. Example: "You are my research partner. You are concise, skeptical of hype, and always cite your sources. You call out when I'm being vague. You do not produce generic listicles."
Hit Save, then Chat with the agent. Ask it something you'd normally ask — see if the personality comes through. If not, iterate. Souls take a few drafts.
You can also edit floor-level and room-level souls from the ✨ Soul button in the Tower's top nav. Floor souls apply to all rooms on your floor; room souls are more specific.
The interview process is how you onboard an agent to your actual life — not just its soul, but the specifics of your context, goals, and preferences. Think of it as HR onboarding, except you're the one being hired.
Open a Chat with your agent and say: "I want to align you to how I work. Ask me 5-7 questions to understand my context, goals, and how I like to be helped."
Tell it about your role, your schedule, your communication style, what you hate in an AI response (bullet lists? verbosity?), what you find useful. Be specific — vague answers produce vague alignment.
Say: "Based on what I told you, write a summary of how you'll approach helping me. I'll tell you if anything's off." This is your calibration check.
Copy the summary into the Settings tab → System Prompt. Now it's persistent — every conversation starts pre-aligned.
Re-run the interview every few months. Your goals change; your agent should too. An agent calibrated to your January self may be subtly wrong by June.
Each floor comes pre-loaded with a set of agents, but you can spawn new ones in any room. Each agent is a specialised worker — the more specific their job, the better they do it.
On your floor in the building view, click [ + ADD ROOM ]. Give it a name and pick its archetype (Researcher, Calendar, Data, Executor, Gmail, Custom). Each archetype has different tools.
Bad: "Agent 1". Good: "Meal Planner", "Minecraft Research Bot", "Story Generator". Your agents are a team — names matter for the Orchestrator when it routes tasks.
In the agent's soul, tell it what it's allowed to do: "You can search the web, read files, and draft emails. You cannot send anything without my explicit approval."
Each agent tracks task history and quality scores. You'll see a score badge on its room card. A dropping score means it needs a better prompt, a harder model, or both.
Your floor has a Chief Orchestrator agent. It's the one that receives your task and delegates to the right room agents. Don't delete it. You're welcome to ignore this warning and find out what happens. (Nothing good.)
Your agents get better the more honest you are with them. Feedback is not just nice to have — it's what separates a good agent from a great one.
After any response, tell the agent if it was wrong, incomplete, or off-tone. Be direct: "That answer was too vague. I needed specific steps." It incorporates feedback immediately.
If you keep correcting the same behaviour, that's a signal to fix the soul. Persistent corrections belong in the system prompt, not the chat window.
Each agent has a quality score badge on its room card. Scores below ~55% mean the agent is consistently missing. Low scores trigger automatic model upgrades — but you can also manually intervene.
In an agent's Memory tab, you can promote important context to the household namespace. This means ALL agents on ALL floors can see it — useful for shared facts like "family dinner is Sunday 6pm."
These limits apply to all agents, all floors, all users. They cannot be overridden by any soul, any prompt, or any request. If you ask an agent to do any of these things, it will refuse. This is intentional.
Everything is stored on the Mac Studio in our house. We use Cloudflare's edge for identity and routing only — your conversations and memories do not live at Cloudflare. The only exception is when a task is routed to Claude (the cloud fallback), which is logged, budget-capped, and shown to you.
Your Google account is your identity key. Never share it. If you suspect someone outside the family has your Google credentials, change your password immediately and tell Cathy so the allowlist can be audited. Access is also protected by Cloudflare's MFA layer.
All agent actions are written to the audit log — who triggered what, when, what model was used, what it cost. Cathy and admins can review this. This isn't surveillance; it's how we catch bugs, review quality, and verify security. If something felt wrong, there's a log entry for it.
The most common questions, answered honestly.