Agent-to-Agent Protocol

Vibe Coding
from Anywhere

Send a WhatsApp message. Your Mac starts coding. The agent asks questions. You answer from bed. Ship features in your pajamas.

Sovereign Brain
A2A Server
Intel Layer
WhatsApp → Orchestrator
live
$ linkbox-orchestrator --port 8080 --a2a 8081
🧠 Sovereign Brain online | 💓 Heartbeat active
🌐 A2A WebSocket ready on :8081
📨 Incoming message from WhatsApp:
A2A :8081
MCP (internal tools)
Message Routing

How a Message Becomes Code

Every message from every platform hits a single front door. From there, the router decides: cloud AI, local Orchestrator, or full agentic coding session.

A2AWebSocket
WhatsApp
Slack
Telegram
Router
Orchestrator
01

Message Arrives

You send a message from WhatsApp, Slack, Telegram, Discord, or SMS.

02

Intent Classified

The intelligent-message-router classifies your intent against the skill manifest.

03

Route Decided

Orchestrator tasks → A2A WebSocket. General queries → Cloud AI. Everything agent goes through A2A.

04

Agent Executes

Your Mac Orchestrator receives the task, reasons, acts, and reports back.

Hyper Symbols

Speak in Symbols

Prefix your message with a hyper symbol for instant, deterministic routing. No AI classification needed. No ambiguity. Just precision.

>>

Sovereign Brain

Route directly to your IDE's coding brain via A2A protocol

>> search for auth logic
>> status
>> ask How does the WebSocket reconnect?
>> diag
!>

Browser Automation

Dispatch a task requiring browser interaction via Computer Use

!> sign up for T-Mobile prepaid
!> book a flight to Miami
!> order groceries from Instacart
/code

IDE Commands

Direct coding session control — goals, files, heartbeats

/code goal Fix the login timeout bug
/code open src/auth/login.swift
/code heartbeat
/code answer Yes, use production DB

No prefix? AI figures it out.

Messages without a hyper symbol are classified by GPT against the skill manifest (60+ intents). This works great for natural conversation, but symbols give you instant, deterministic control when you know exactly what you want.

Architecture

Five Layers of Intelligence

The Orchestrator isn't one monolith — it's a stack of specialized layers, each doing one thing exceptionally well.

Sovereign Brain

Coding Session Manager

A2A layer
  • Multi-turn conversation thread
  • Autonomous task heartbeat (5s pulse)
  • Task phase management (plan → code → test)
  • User question/answer dialogue
  • Stall detection (30s timeout)

A2A Server

Agent-to-Agent Protocol

port 8081
  • WebSocket bidirectional dialogue
  • Task request / accept / complete lifecycle
  • Safety confirmations for risky actions
  • Progress updates streamed in real-time
  • Question/answer with timeout handling

Intel Layer

Code Intelligence

A2A + REST
  • Semantic vector search across codebase
  • Real-time diagnostics (errors, warnings)
  • LLM-powered code completion
  • Context pruning for token budgets
  • AST-aware repo map generation

Omni-Channel Gateway

Universal Platform Bridge

Edge Function
  • WhatsApp, Slack, Telegram, Discord, SMS
  • Alexa, Google Home, Gemini Live voice agents
  • Hyper-symbol prefix routing (>>, !>, /code)
  • All agent tasks routed via A2A — one artery
  • Falls back to cloud AI when Orchestrator offline

BennyBoy Engine

Browser Automation

Computer Use
  • Gemini Computer Use via Vertex AI
  • High-fidelity CDP browser control
  • Multi-step agentic web tasks
  • Screenshot reasoning loop
  • Google account pre-selection

MCP Server

Model Context Protocol

IPC
  • File system operations (read/write/search)
  • Terminal command execution
  • Memory store (remember/recall)
  • Entry search via RAG
  • Git operations & GitHub integration

Two Arteries, One Brain

Not 1,000 small connections — two clean, purpose-built communication channels. A2A for every agent conversation. MCP for every tool action.

A2A — The External Artery

port 8081 • WebSocket • bidirectional

Every agent speaks A2A. WhatsApp, Slack, Alexa, Google Home, Gemini Live, the web IDE — they all connect to the Orchestrator through one universal WebSocket protocol. Tasks, questions, progress, safety checks — all flow through this single artery.

Message Lifecycle
a2a_task_request // "Fix the auth bug"
a2a_progress // 40% "Analyzing codebase..."
a2a_question // "Which database?"
a2a_user_response // "Use production DB"
a2a_complete // "Bug fixed, 3 files changed"
WhatsAppSlackAlexaGoogle HomeGemini LiveWeb IDE

MCP — The Internal Artery

Model Context Protocol • tool execution

Agents don't touch tools directly. When the Orchestrator needs to read a file, run a terminal command, search your knowledge base, or push to GitHub — it invokes MCP tools internally. This keeps the tool layer isolated, secure, and composable.

Available Tools
read_file / write_file / list_directory
run_terminal_command
search_entries (RAG knowledge base)
remember / recall (persistent memory)
git_status / git_commit / github_pr

Why not HTTP? HTTP (port 8080) exists only for the Link Code Pro web IDE bridge — a user-facing UI, not an agent. All agent-to-Orchestrator communication goes through A2A.

Design Principle: Two arteries, not 1,000 connections

Instead of building separate HTTP endpoints for every feature (status, search, diagnostics, goals, files...), we funnel everything through two clean channels. A2A handles the conversation. MCP handles the tools. No fragmentation. No confusion.

Link Code Pro

A Full IDE in Your Browser

Monaco editor. File tree. Terminal. Git. AI chat with 5 modes. All connected to your Mac via the Orchestrator bridge.

Monaco Editor

Full VS Code editing experience with syntax highlighting, IntelliSense, and multi-tab support

Integrated Terminal

Run shell commands on your Mac remotely via the Orchestrator bridge

Git Integration

Stage, commit, push, diff — full version control without leaving the browser

AI Chat (5 Modes)

Code, Architect, Ask, Debug, Edit — each tuned for a different workflow

Semantic Code Search

Vector-powered search that understands meaning, not just keywords

Self-Healing Loop

AI auto-fixes diagnostics: generates patch → applies → re-checks until clean

Context Pruning

Token-aware context window: target file + imports + semantic chunks + repo map

AI Code Completion

Ghost text completions powered by your chosen LLM provider

Live Diagnostics

Real-time error and warning feed from the Orchestrator Intel Layer

Code from Everywhere

Every messaging platform becomes a coding terminal. Same commands, same brain, same results.

WhatsApp
Slack
Telegram
Discord
SMS

Voice Agents — All via A2A

Amazon Alexa
Google Home
Gemini Live
Link Code Pro

Ready to Vibe Code?

Install the Orchestrator on your Mac. Connect your messaging apps. Start shipping from anywhere.