AI Operating SystemClaude CodeMulti-AgentMCP ProtocolBusiness AutomationSolo Build

SmithClaude

A full AI operating system built on Claude Code that runs my business — from LinkedIn prospecting and content creation to calendar management, document generation, and cross-channel communication. Features a team of AI employees (Jamie for sales, Fiona for content), 30+ custom scripts, persistent vector memory, and integrations with Notion, Gmail, Google Calendar, Supabase, and Telegram.

The Problem

Running a solo AI consultancy means wearing every hat — sales, marketing, content, admin, client work. I needed an AI system that didn't just answer questions but actively ran business operations: prospecting leads on LinkedIn, writing personalised DM sequences, creating content, managing calendar and email, generating documents, and maintaining context across days and channels. Existing AI assistants were either too generic (ChatGPT) or too narrow (single-purpose automations). What I needed was a unified AI operating system with specialised agents, persistent memory, and the ability to take real action across every business tool.

Key Metrics

154 leads in one session

Sales pipeline

10 lead magnets in 30 mins

Content velocity

30+ custom scripts

Capabilities built

3 AI employees

Team size

My Role & Approach

I built this on Claude Code as the backbone — leveraging its native tool use, MCP protocol support, and persistent session architecture. I created a skill-based system where each business function is handled by a specialised AI employee: Jamie (LinkedIn Sales Agent) handles prospecting and DM sequences, Fiona (Content Engine) writes LinkedIn posts and lead magnets. I connected everything through MCP servers: Notion for project management and documents, Google Workspace for email and calendar, Supabase for CRM and vector memory, Telegram for mobile access. I built 30+ custom scripts for capabilities Claude Code doesn't have natively — browser automation, image generation, video generation, voice transcription, usage tracking. I implemented cross-channel memory using Supabase pgvector so context persists across terminal, Telegram, and web sessions, and added a security layer for Telegram with injection detection, rate limiting, and anomaly detection.

Tools & Tech

Other

Claude CodeClaude Opus 4MCP ProtocolSupabase (pgvector)Notion APIGoogle WorkspaceTelegram Bot APIPythonTypeScriptTailwind CSS

The Build Process

Architecture & Foundation

I designed the skill-based agent architecture where each business function maps to a specialised skill file. Set up CLAUDE.md as the system's source of truth — containing all protocols, tool references, and behavioral rules. Connected MCP servers for Notion, Google Workspace, Supabase, and Telegram.

Memory & Continuity

I built a persistent memory system using Supabase pgvector with local embedding generation (all-MiniLM-L6-v2). Implemented session handoff protocols so context carries across conversations. Created cross-channel memory unification — facts learned in Telegram are available in terminal and vice versa.

AI Employees & Skills

I created Jamie (LinkedIn Sales Agent) with Apify integration for lead sourcing, qualification scoring, and personalised 4-message DM sequence writing. Built Fiona (Content Engine) for LinkedIn thought leadership posts. Developed the auto-persona skill for dynamic expert role selection. Each employee has clear scope, tools, and escalation rules.

Capabilities & Scripts

I built 30+ Python scripts extending Claude Code's native abilities: browser automation (Playwright), image and video generation (kie.ai), voice transcription (Whisper), security scanning, usage tracking, Supabase REST client, Ollama fallback for rate limits, and session management. Each script follows a self-upgrade protocol — when a capability is missing, the system proposes and implements it.

Learnings

“The skill-based architecture was the breakthrough — instead of one monolithic AI trying to do everything, specialised agents with clear scope and escalation rules produce dramatically better results. Memory is still the hardest problem: vector search works well for factual recall but struggles with nuanced context. The self-upgrade protocol changed the dynamic from 'it can't do that' to 'it can build that' — every limitation became a feature request. Building on Claude Code rather than from scratch meant the core reasoning, tool use, and safety features were already world-class, letting me focus entirely on business value. The biggest lesson: an AI operating system isn't about the AI — it's about the system design around it.”