GoGrab Platform — Current Architecture Report

Date: February 26, 2026 Session: 148 (gspace1) Research Method: MCP Research Protocol (mcp-research.md) — Phase 1-4 complete Calibrated Confidence: 0.92 (research), 0.54 (architecture decision domain — historical success rate warning)

Executive Summary

The GoGrab AI Platform is a multi-layered, multi-agent system with 10+ deployed components spanning 4 backend APIs, 5 client surfaces (web, Chrome extension, VSCode extension, desktop, mobile), an MCP coordination server, and 238+ NX libraries. The AI orchestration layer (Vitto) is production-grade with a 4-phase engine, A2A protocol compliance, computer use infrastructure, and a learning pyramid tracking 563 patterns.

Critical gap identified: The platform building capabilities — the libraries that generate, edit, and compose modules/pages/flows — remain fragmented with significant dead code and poor integration between generation (vibe-universe) and editing tools (universe-canvas, page-builder, flow-engine, layout-builder, module-editor-lib, etc.). This is documented separately in a follow-up research.

1. Backend Infrastructure

1.1 Four-API Architecture

1.2 Vitto Engine (backend-vitto-api)

206 TypeScript files across 17 modules:

Key services by size:

• vitto-engine.service.ts (1,804 lines) — task submission, queue management, SSE streams
• phase-executor.service.ts (1,000+ lines) — 4-phase workflow execution
• module-generation.service.ts (600+ lines) — code generation pipeline
• mind.service.ts (500+ lines) — memory/context management
• a2a.service.ts (200 lines) — A2A protocol translation

Test coverage: 36 spec files including 5 A2A protocol test suites (59 tests).

1.3 Backend AI API (backend-ai-api)

325 TypeScript files across 26 modules:

2. Computer Use Infrastructure

Location: APIs/backend-ai-api/src/app/infrastructure/ Scope: 74 files (30 services + 6 sandbox providers + 30+ skills)

Architecture

SandboxProviderFactory (env-based selection)
├─ SelfHostedProvider    (SANDBOX_SERVER_URL → Local Node.js)
├─ E2BCloudProvider      (E2B_API_KEY → Firecracker microVMs)
├─ PlaywrightProvider    (fallback → Browser automation)
└─ RobotjsProvider       (@jitsi/robotjs → Desktop control)

ComputerUseLoopService (1,269 lines)
├─ Observe: Screenshot + OCR
├─ Think:   Anthropic Claude (computer_20251124)
├─ Act:     Execute via selected sandbox provider
└─ Guards:  25 iterations, 100K tokens, 5min timeout, cost tracking

• E2B Desktop: Firecracker microVMs, Chrome pre-warmed, Firefox killed, 30min idle TTL
• Mid-session reply: POST /sessions/:id/reply queues user messages
• Service profiles: 9 static (Gmail, Calendar, GitHub, etc.)
• Skills: 30+ implementations (browser, SSH, Docker, Git, desktop)

3. A2A Protocol (Agent-to-Agent)

Location: APIs/backend-vitto-api/src/app/a2a/ Status: Complete with full test coverage (59 tests across 5 suites)

• Bridges Google A2A protocol to internal VittoTask system
• MongoDB persistence with 7-day hot cache
• Status mapping: internal states → A2A standard (working/completed/failed/canceled)
• SSE streaming for real-time task updates

4. MCP Server (Model Context Protocol)

Location: MCP/mcp-gograb-ai/ Scope: 132 TypeScript files, 64+ tool handlers

Tool Categories

• Core workflow: preflight, log_decision, outcome
• Intelligence: recall, simulate, reflect, diagnose, calibrate (21 actions)
• Session management: get_info, set, summary, validate, lifecycle
• Code operations: review, test_gen, validate, analyze
• Agent collaboration: handoff, orchestrate, message, broadcast, rpc, discover
• Research: web_search, docs_search
• Browser/Desktop: navigate, click, fill, screenshot, execute
• Git: SAGP-enforced commit, stash, merge_check, recover

Key Features

• Sentinel reconnect (10 fast retries → 60s intervals indefinitely)
• Offline handoff sync via REST with 2x exponential backoff
• Platform detection: Cursor, Claude Code, Claude Desktop, VSCode
• Notification injection on tool responses (pull-only, 60s check)
• Connects exclusively to backend-ai-api (port 3001)

5. Client Applications

5.1 Web Application (Angular 20)

Location: apps/gograb-ai/ Pages: panel, vitto-chat, dashboard, observatory, home, extension-panel, desktop-auth, desktop-welcome

• AI Panel with streaming chat, drag-drop "Add to Canvas"
• Observatory dashboard (30s auto-refresh, agent monitor + system dashboard tabs)
• EmbeddedAdapterService for iframe bridge support
• Session persistence via MongoDB (not autoFlashState — race conditions confirmed)

5.2 Chrome Extension (Vitto)

Location: apps/vitto-chrome-extension/ Version: 0.2.9 | Manifest V3

• Side panel (Ctrl+Shift+G / Cmd+Shift+G), popup (Alt+G)
• Content script on GoGrab domains
• Bridge adapter (postMessage) for embedded iframe communication
• Auth via URL params in extension context (not postMessage — unreliable)

5.3 VSCode Extension (Vitto Agents)

Location: apps/vitto-vscode-extension/ Version: 0.9.7 | 35 TypeScript files

• 13 commands (status, handoffs, preflight, logDecision, findSimilar, reviewCode, etc.)
• Full MCP connection + tool routing
• Native VSCode notification modals for handoffs
• Webview panels for inline UI

5.4 Desktop App (Tauri)

Location: apps/vitto-desktop/ Build: Tauri (Rust backend + Angular frontend)

• Global hotkey: Cmd+Ctrl+V (macOS)
• CI/CD passing all 4 platforms (macOS arm64/x64, Windows x64, Linux x64)
• Auto-updates via Tauri updater
• Code signing with ~/.tauri/vitto.key

5.5 Mobile App (Capacitor)

Location: apps/vitto-mobile/ App ID: space.gograb.vitto.mobile

• Configured, uses same Angular app via webDir
• CapacitorBridgeService mirrors TauriBridgeService pattern
• Auth guard covers: isDesktop || isMobile || isEmbedded
• Needs iOS/Android signing pipeline

6. Learning Pyramid

7. Branding & Domain Mapping

8. Architecture Patterns

1. Strategy/Factory pattern — Used consistently for sandbox providers, LLM providers, auth bridges
2. 4-phase workflow — ANALYZE → PLAN → EXECUTE → VERIFY (Vitto engine + gmodules1)
3. Build UI, Buy Engine — Custom UX components, wrapped external parsers (marked, PrismJS)
4. MCP-centric coordination — All agent types coordinate through MCP, not direct API calls
5. Offline-first resilience — REST sync, local queuing, sentinel reconnection

9. Risk Assessment

High Priority

• 51.2% architecture decision success rate — Nearly half of decisions result in partial/full failure. Audit the 24 failed outcomes for root causes.
• Platform builder fragmentation — Generation (vibe-universe) and editing tools (universe-canvas, page-builder, flow-engine, etc.) are poorly integrated with significant dead code. (See dedicated follow-up research.)

Medium Priority

• autoFlashState race conditions — Confirmed unresolvable. Always use MongoDB for critical data.
• A2A hot cache 7-day limit — No validation on restart, cold starts lose task state.
• LLM cost alerting absent — Observatory tracks but doesn't alert on threshold breaches.

Low Priority

• Mobile deployment pipeline — Capacitor configured but no signing pipeline.
• Backend trace/pivot detection offline — Intelligence tools partially unavailable when backend disconnected.

10. Port Reference

11. Key File Paths

Generated via MCP Research Protocol. Outcomes: 162 validated, 563 patterns, 64+ MCP tools operational.