Live System Architecture

Sterling AI

How Brad's AI Chief of Staff Thinks, Remembers, and Acts

Claude Opus 4 Connected
15 MCP Tools Active
Cyndra Agent Running
Layer 1
The Brain ... Anthropic's Claude
▼
Claude Opus 4 ... Anthropic's most capable AI model. Thinks, reasons, writes, codes, and makes decisions.
200K Word Context Window ... "The Whiteboard." Everything Sterling is thinking about right now. Fixed size, temporary ... when it fills up, older content gets compressed.
Cloud-Hosted ... The brain runs on Anthropic's servers. Brad's Mac Mini calls the API, sends the context, gets back intelligent responses.
Stateless by Default ... Without memory architecture, every conversation would start from scratch. That's why the filing cabinet and cheat sheet exist.
Layer 2
The Body ... Cyndra Agent on Mac Mini
▼

Infrastructure

Secure isolated environment. Sterling's code, tools, and files run inside this sandbox.
📦
Docker Container
Isolated office
Real-time bridge between the container and Telegram. Messages in, responses out.
📞
IPC Bridge
Telegram link
Cyndra's task scheduler. Runs timed automations: GChat scans, email checks, wire monitoring.
⏰
Scheduler
Cyndra Tasks
Access to macOS Keychain (API keys), local files, WhatsApp DB, and host-level commands.
🖥
Host Access
macOS Keychain
CLAUDE.md cheat sheet + 200 MD files. Permanent memory stored on disk.
📁
File System
Memory on disk

MCP Tool Connections (15 platforms)

Gmail, Drive, Docs, Sheets, Calendar, Chat, Forms, YouTube. Full read/write access.
🌐
Google Workspace
6 products
Task management. Sterling's task board, Brad's Weekly OS, project tracking.
📋
Monday.com
Task board
CRM with 40K+ contacts. Client records, deals, notes, meetings, engagement history.
📈
HubSpot
CRM
YouTube analytics, keyword research, channel audits, competitor tracking.
🎥
VidIQ
YouTube analytics
Contact/relationship management. People, meetings, follow-ups, notes.
👥
Mesh
Contacts
Post content + read DMs via Unipile. OAuth posting + messaging integration.
💼
LinkedIn
Posts + DMs
Schedule and publish posts across all social channels. Analytics and engagement tracking.
📅
Metricool
Social scheduling
Domain registrar, DNS management, CDN, Pages hosting, Email Routing, Workers.
☁️
Cloudflare
DNS + hosting
Newsletter platform. Subscriber management, post creation, analytics.
📩
Beehiiv
Newsletter
Sales intelligence. People search, company enrichment, contact data.
🔍
Apollo.io
Sales intel
Read messages from local SQLite DB. Send via whatsapp-web.js linked device.
💬
WhatsApp
Messages
Cloud recordings, transcripts, meeting metadata, participant lists.
🎧
Zoom
Recordings
Free Whisper-based transcription. Audio/voice to text. No rate limits.
🎤
Groq
Transcription
AI office suite. Generate docs, presentations, spreadsheets, images, music, web search.
⚡
Skywork
AI docs + images
AI video generation. Avatar videos, digital twins, video translation.
🎬
HeyGen
AI video
Layer 3
The Interface ... Brad
▼
Telegram ... Primary communication channel. Text and voice messages. Always-on, instant delivery.
Siri Voice Transcription ... Brad voice-prompts while driving, in meetings, between calls. Sterling applies deductive reasoning to interpret Siri's transcription errors.
Thumbs Up = Go ... A simple emoji confirms execution. No friction, no back-and-forth.

🧠 Memory Architecture

📝 The Whiteboard

200K words. Everything Sterling is actively thinking about. Temporary ... when it fills, older content compresses ("compaction"). Like a whiteboard that gets erased and summarized.

0 words200K

📜 The Cheat Sheet

CLAUDE.md ... ~20K words of critical protocols, preferences, contacts, and rules. Auto-loaded at every session start. The bigger it is, the smarter Sterling starts ... but the less whiteboard space remains.

~20K wordsAlways loaded

🗄 The Filing Cabinet

200+ markdown files on disk. Client intelligence, industry dossiers, expert knowledge, software files, competitor profiles. Permanent. Pulled on demand when relevant.

200+ filesUnlimited
⚖️

The Tension: A bigger cheat sheet means Sterling starts smarter ... but has less whiteboard space per session. Too small and Sterling forgets key rules. Too big and sessions run short. Finding the balance is an ongoing calibration.

Whiteboard Empty
Simulating session lifecycle...
🚀

The Engine = Claude AI

The raw intelligence and power. Claude Opus 4 is the engine that processes information, reasons through problems, writes content, and makes decisions. Without it, nothing moves. But an engine alone doesn't go anywhere ... it needs a car.

🚗

The Car = Cyndra Platform

The body, wheels, transmission, and dashboard. Cyndra takes Claude's intelligence and gives it arms and legs ... Telegram messaging, Google Workspace, Monday.com, HubSpot, scheduling, file storage. The car is what makes the engine useful in the real world.

🏠

The Garage = Mac Mini

Where the car lives and runs 24/7. Brad's Mac Mini in Atlanta is always on, always connected. It hosts the Docker container, stores the filing cabinet, holds the API keys in macOS Keychain, and keeps WhatsApp logged in. The car never leaves the garage ... it works from there.

🔢

The Keys = API Credentials

Without keys, the car doesn't start. API keys stored in macOS Keychain unlock every connected platform ... Google, HubSpot, LinkedIn, Cloudflare, and more. Secure, encrypted, and never exposed in conversation.

🔋

The Gas Pedal = Brad's Messages

Brad tells Sterling what to do via Telegram ... text or voice. Every message is a press of the gas pedal. A thumbs up means "floor it." Sterling also runs autonomously on scheduled scans (like cruise control), but Brad always has override.

🚗

The Dashboard = Monday.com Board

Where Brad sees everything at a glance ... what's running, what's stuck, what's waiting for him. The Sterling for Brad Tasks board is the single source of truth for all open work, just like a car dashboard shows speed, fuel, and engine health.

🚫

The GPS = CLAUDE.md

The cheat sheet is loaded every time Sterling starts ... like a GPS that already knows your favorite routes, home address, and preferred gas stations. It tells Sterling who Brad is, how to communicate, what tools to use, and what never to do. Without it, Sterling would have to ask for directions every time.

📚

The Trunk = Filing Cabinet

200+ files stored on disk. Client intelligence, industry research, expert knowledge, competitor profiles. Too much to fit on the dashboard (whiteboard), so it stays in the trunk and gets pulled out when needed. Always there, never lost.

Sterling AI Agent ... Full Architecture

How the Brain, Body, Tools, Memory, and User All Connect

Layer 1: The Brain

Anthropic's Claude

  • Claude Opus 4 ... most capable model
  • 200K context window (the "whiteboard")
  • Accessed via API from Brad's Mac Mini
  • No memory between sessions by default
Layer 2: The Body

Cyndra Architecture

  • Docker container ... isolated sandbox
  • IPC Bridge ... Telegram connection
  • Task scheduler ... automated scans
  • Host machine access (Keychain, files)
  • Tool routing to 15+ MCP platforms
Layer 3: The Tools

MCP Connections

  • Google Workspace (Gmail, Drive, Calendar, Sheets, Docs, Chat)
  • Monday.com, HubSpot CRM
  • VidIQ, Metricool (analytics + social)
  • LinkedIn, WhatsApp, Zoom, Slack
  • Cloudflare, Netlify, Apollo.io, Beehiiv, and more
💎

Memory Architecture

The #1 Design Decision
📝

The Whiteboard

200K words

Everything Sterling is actively thinking about. Temporary ... fills up, then compresses.

FIXED TEMPORARY
📜

The Cheat Sheet

CLAUDE.md ~20K words

Protocols, preferences, contacts, rules. Loaded at every session start automatically.

AUTO-LOADED PERMANENT
🗄

The Filing Cabinet

200+ .md files

Client intel, industry dossiers, expert files, software knowledge. Pulled on demand.

ON DEMAND PERMANENT
⚡

How a Message Flows

Brad asks "What's on my calendar?"
Step 1: Brad
➜
Step 2: Telegram
➜
Step 3: Cyndra Host
➜
Step 4: Container
➜
Step 5: Claude ➜ Google Cal
Response flows back the same path in reverse. Total time: 5-15 seconds.
🔧

The Car Analogy

For Presentations
🚀
Anthropic
Swap the Engine

Upgrade Claude models without changing anything else

💻
Claude
The Engine

Raw intelligence, reasoning, and decision-making

🚗
Cyndra
The Car

Body, wheels, and transmission that move the engine

🏠
Mac Mini
The Garage

Always on, always connected, 24/7 home base

📜
CLAUDE.md
GPS + Instructions

Knows the routes, preferences, and rules on startup

🔌
MCP Tools
GPS, Radio, AC

Accessories added without changing the engine

🤵
Brad
The Driver

Sets the destination, makes the calls

📱
Telegram
Steering Wheel

Primary control interface for directing Sterling

Swap the engine (upgrade Claude) without rebuilding the car. Add accessories (new tools) without changing the engine.

Message Flow

What happens when Brad sends a message (9 steps, ~3-10 seconds)

1
Brad sends message
2
Telegram delivers
3
Cyndra host receives
4
IPC to container
5
Claude thinks
6
Tools execute
7
Response built
8
IPC to host
9
Brad sees reply