System Architecture

The Stack

Claude Code + MCP Servers + Sub-Agents + Memory + Skills

Full System View

Architecture Diagram

flowchart TB
  subgraph Operator["Operator Layer"]
    User["Colin / Human Operator"]
    Voice["Voice Announcements<br/>Elisabeth · Finn"]
  end
  subgraph Orchestration["Orchestration Layer"]
    CC["Claude Code CLI"]
    Skills["Skills Engine<br/>12-Phase Coding · TDD · PDF Creator"]
    Memory["Persistent Memory<br/>MEMORY.md + file-based"]
    Tasks["Task Tracking<br/>Progressive completion"]
  end
  subgraph MCP["MCP Server Layer"]
    Chrome["Chrome Browser<br/>Scraping + Testing"]
    Linear["Linear<br/>Project Management"]
    GitHub["GitHub<br/>Version Control"]
    TTS["Edge-TTS<br/>Voice Synthesis"]
    Sheets["Google Sheets"]
    Gemini["Gemini Chat<br/>Cross-validation"]
    OpenAI["OpenAI Chat<br/>Peer review"]
  end
  subgraph Agents["Agent Layer"]
    Sub1["Research Agents ×6"]
    Sub2["Scraper Agents ×10"]
    Sub3["Analysis Agents ×4"]
    Sub4["Dev Agents ×2"]
  end
  subgraph Output["Output Layer"]
    Site["Astro Website<br/>23 pages"]
    Report["Analysis Report<br/>96KB · 16 diagrams"]
    Docs["Handover Docs<br/>CLAUDE.md · DEPLOY.md"]
  end
  User <--> CC
  CC --> Skills
  CC --> Memory
  CC --> Tasks
  CC <--> MCP
  CC --> Agents
  Voice --> User
  Agents --> Output
  MCP --> Agents
        

Layer by Layer

Key Components

Orchestration

Claude Code CLI

The orchestrator runtime. Manages context windows, spawns sub-agents, enforces governance rules, and coordinates the entire pipeline from a single terminal session.

MCP Layer

Model Context Protocol

Standardised tool interface. Each MCP server exposes capabilities — browser, project management, voice, spreadsheets — that agents can call via a uniform JSON-RPC protocol.

Skills

Skills Engine

Reusable skill definitions (markdown files) encoding domain knowledge. Examples: bacon-evolutionary-coding (12-phase methodology), pdf-creator, test-guard.

Memory

Persistent Memory

File-based memory system surviving across sessions. Four types: user preferences, feedback loops, project state, and reference knowledge. Context never gets lost between agent lifetimes.

Git

Git Worktrees

Parallel working directories sharing one repository. Multiple agents can work on different aspects simultaneously without merge conflicts or blocking each other.

Agents

Agent Fleet

22 specialised sub-agents — scrapers, researchers, analysts, developers — each spawned with full context inheritance. They run in parallel, report back, and are governed by the same SA rules.

Protocol Design

Why MCP Matters

Standardised Interface

Each MCP server is a standardised JSON-RPC interface. Chrome browsers, Linear project management, Edge-TTS voice, Google Sheets — all accessed through the same protocol.

Extensible by Design

Adding new capabilities means adding a new MCP server — not modifying the framework. Need Slack integration? Deploy a Slack MCP server. The orchestrator discovers it automatically.

Real Browser Access

The Chrome browser MCP server lets agents literally browse the web — click buttons, fill forms, take screenshots, scrape content. Not simulated. Real browser automation.

7 Servers in This Project

Algorithmix used 7 MCP servers simultaneously: Chrome (scraping), Linear (tracking), GitHub (version control), Edge-TTS (voice), Google Sheets (data), Gemini and OpenAI (peer review).

Cross-Validation

Multi-Model Peer Review

BACON-AI does not just use Claude — it cross-validates with Gemini and GPT to catch model-specific blind spots.

C

Claude

Primary orchestrator

Runs the pipeline, generates content, spawns agents, writes code

G

GPT

Peer reviewer

Challenges assumptions, reviews analysis quality, provides alternative angles

G

Gemini

Cross-validator

Root-cause analysis, debugging triage, validates solutions from a different perspective

The Triage Protocol

1

After 3 failed fix attempts, mandatory triage is triggered

2

Claude shares the complete error context with both GPT and Gemini

3

All three models provide independent root-cause analyses

4

Solutions are synthesised, tested, and cross-validated before applying

“Three AIs disagreeing is more valuable than one AI agreeing with itself.”