AI Stack

Multi-World AI Stack — 5 layers, 30+ tools, one coherent operating system

Multi-World AI Stack
Every layer of the AI Shop engine
Board → Compute → Agents → Domains → Integration Patterns. This stack map defines how Claw-Kanban, Perplexity Computer, 30+ AI agents, and four application domains connect into a single event-driven platform. Use it to understand where each tool lives, what it owns, and how work flows between layers.
5
Layers
30+
Tools
4
Domains
3
Pipelines
Part of AgentOS
This AI Stack describes the tools and compute inside AgentOS. The full 7-layer operating system — Interface, Coordination, Planner, Intelligence, Execution, Memory, and Culture — is defined here:
AgentOS →
Architecture Diagram
Layer 1 · Board & Orchestration
Claw-Kanban · Trello / Jira · Notion / Obsidian
↓
Layer 2 · Core Compute
Perplexity Computer · Vane (Private) · GPU Clusters · Edge Nodes
↙↓
General LLMs
Claude · Gemini · Codex
Reasoning / Autonomous
AutoGPT · LangChain · HuggingFace
↓
Multimodal
DALL·E · Runway · LLaVA
Knowledge & Research
Vane · Elicit · Wolfram · Pinecone
Domain-Specific
Harvey · MedPaLM · AlphaSense
↓
Simulation / Virtual Worlds
Automation & Robotics
Edge / IoT AI
Self-Updating KB
Key Opportunities from This Stack
Orchestration
Event-Driven Agent Pipelines
Connect Claw-Kanban task completion → Perplexity Computer trigger → agent execution chain. Each product becomes a node in an automated workflow, not a manual job.
Compute
Ollama-First, Cloud Fallback
Local GPU via Ollama keeps inference cost at 92% margin. Cloud (Claude / Gemini / Codex) is reserved for tasks that exceed local capacity — triggered by confidence score or complexity flag.
Agents
Specialist Agent Network
Each AI agent owns a specific task class — no agent tries to do everything. LangChain pipelines route tasks to the right specialist: research → summarization → code → visualization.
Knowledge
RAG-Powered Products
Pinecone / Weaviate vector stores power ChatDesk's KB, SEOAgent's research memory, and AgentBenchmark's challenge archive. Knowledge persists across agent sessions via SharedContext MCP.
Domains
Multi-World Expansion
Simulation (Unity ML-Agents), RPA automation (UiPath), Edge AI (TinyML), and a self-updating knowledge base are post-Phase-3 expansion vectors that grow the platform beyond SaaS.
Multimodal
Content & Media Automation
SEOAgent can use DALL·E for article images. Runway ML enables video demos for AgentArena. LLaVA adds vision capabilities for screenshot-based testing in ShipIt's review loop.
1
Board / Project Management
Task routing & knowledge linking
Claw-Kanban Claw-Empire (CEO) Trello connector Jira connector Notion Obsidian

Claw-Kanban is the primary task dispatch layer. Claw-Empire acts as the CEO-level orchestrator — it receives strategic directives via Slack ($ prefix) and fans them into coordinated multi-agent missions. Trello and Jira connectors let automated agent progress updates flow into your existing project management workflow without manual sync. Notion and Obsidian serve as the long-form knowledge layer — agent-generated documentation, architectural decisions, and research summaries are written there for human review.

2
Core Compute
Central AI execution environment
Perplexity Computer Vane (Private Research) Ollama (local inference) GPU Cluster / Cloud Instance Edge Nodes (IoT / Simulation)

Perplexity Computer is the execution backbone — it runs agents, deploys artifacts, manages the workspace, and coordinates tool calls. Vane is the private research engine — a self-hosted AI answering engine (SearxNG + local LLMs) that analyzes sensitive business data (financials, contracts, agreements) without any data leaving your infrastructure. Ollama handles the majority of inference locally (qwen3-coder:30b primary, falling back to cloud APIs only when confidence is low or context exceeds local capacity). GPU clusters are an optional upgrade path for multimodal tasks that need heavy compute. Edge nodes are a post-Phase-3 expansion for IoT-based AI products and distributed simulation work.

3
AI Agents
5 categories · 20+ specialized tools
A General-Purpose LLMs
The three primary coding agents that work hand-in-hand. Each has a distinct specialty — see AGENTS.md for full role division and routing directives.
Claude Code Gemini CLI Codex

Claude → architecture & security review
Gemini → research, content, long-context
Codex → algorithms, tests, UI components

B Specialized Reasoning / Autonomous
Self-directed agents and pipeline orchestrators. Useful for multi-step autonomous tasks, API chain orchestration, and custom model fine-tuning.
AutoGPT BabyAGI LangChain HuggingFace Transformers

AI Shop opportunity: LangChain as the backbone of AgentOps' multi-step reasoning pipeline. AutoGPT for fully autonomous background research jobs in SEOAgent.

C Multimodal Agents
Image generation, video manipulation, and vision-language reasoning. Adds visual output capability to otherwise text-only pipelines.
DALL·E 3 MidJourney Stable Diffusion Runway ML LLaVA

AI Shop opportunity: DALL·E for SEOAgent article images, Runway for AgentArena demo videos, LLaVA for screenshot-based UI test validation in ShipIt.

D Knowledge & Research Agents
Literature summarization, math/computation, and vector retrieval. The memory and research backbone — powers RAG across all 7 products. Dual-researcher pattern: Perplexity for public web, Vane for private data.
Vane (Private Research) SearxNG Elicit Genei Wolfram Alpha GPT Code Interpreter Pinecone Weaviate

AI Shop opportunity: Vane for private financial analysis and contract review across ChatDesk and InvoiceAgent. Pinecone / Weaviate as vector store for ChatDesk KB ingestion, SEOAgent research memory, and AgentBenchmark challenge archive. Wolfram for financial computation in InvoiceAgent.

E Domain-Specific Agents
Vertical-specialist AI for legal, medical, and financial tasks. Post-Phase-3 expansion — these unlock enterprise-grade offerings in high-margin professional verticals.
⚖️ Legal
Harvey.ai Casetext

Contract analysis, legal document summarization. Potential add-on for enterprise ChatDesk clients in US market.

🏥 Medical
MedPaLM Babylon Health

Clinical reasoning and patient summaries. High regulatory overhead — post-Phase-3 only, requires compliance review before building.

📈 Finance
AlphaSense Bloomberg GPT

Financial insight extraction, earnings summaries. Natural fit for InvoiceAgent upsell: automated financial reporting layer for Nigeria/CR SMBs.

4
Application Domains
Simulation · Automation · Edge · Knowledge
🎮
Simulation & Virtual Worlds
Unity ML-Agents and Unreal AI enable AI agents to be tested inside virtual environments before deployment to production. Key for AgentArena's challenge infrastructure — agents compete in sandboxed simulations before real-world code execution.
Unity ML-Agents Unreal AI
🤖
Automation & Robotics
UiPath and RPA agents handle repetitive task automation — form filling, data extraction, browser actions. Direct integration path for AgentOps' enterprise tier: replace manual workflows with agent-executed RPA for larger SMB clients.
UiPath RPA Agents
📡
Edge / IoT AI
TinyML and on-device inference push AI to the edge — enabling distributed multi-world operations. Post-Phase-3 expansion: IoT product monitoring, sensor-triggered agent workflows, offline-capable InvoiceAgent for markets with unreliable connectivity (Nigeria).
TinyML On-device inference
🧠
Self-Updating Knowledge Base
An automatic research + indexing agent continuously feeds the platform's knowledge layer. Gemini CLI runs scheduled research runs, stores findings in Pinecone, and surfaces them to all other agents via SharedContext MCP — keeping competitive intelligence, market data, and API documentation perpetually current.
Scheduled Gemini runs Pinecone / Weaviate SharedContext MCP
5
Workflow & Integration Patterns
How layers connect in practice
Meta-Agent Coordination
Agent Primary Role Receives From Hands Off To
Claude Code Strategy, architecture, security review Claw-Kanban tasks, Gemini research output Codex (implementation), Gemini (docs)
Gemini CLI Research, content, long-context synthesis Claw-Kanban tasks, Vector DB queries Claude (architecture input), SharedContext
Codex Algorithm design, test generation, UI Claude design specs, LangChain chains Claude (PR review), test results → SharedContext
LangChain Multi-step pipeline orchestration AgentOps Slack trigger Codex (code), Claude (review), Gemini (docs)
Vector DB Knowledge retrieval (RAG) Gemini (write), all agents (read) ChatDesk, SEOAgent, AgentBenchmark context
Event-Driven Pipelines
Pipeline 1 — Research → Content → Publish (SEOAgent)
Trigger Claw-Kanban New keyword task dispatched
→
Research Gemini CLI SERPs + Elicit + vector store lookup
→
Image Gen DALL·E 3 Article hero + inline images
→
Draft Gemini CLI Long-form article + SEO score
→
Publish Codex CMS API push + index update
Pipeline 2 — Task → Build → Review → PR (AgentOps / ShipIt)
Trigger Slack /task User submits task description
→
Architect Claude Code System design + spec.md
→
Tests Codex Write failing tests first
→
Build Codex Implement until 100% pass
→
Review Claude Code Security + quality PR review
Pipeline 3 — Support Query → KB Lookup → Response → Escalation (ChatDesk)
Input WhatsApp / Web Customer query arrives
→
Retrieve Pinecone RAG Vector similarity search on KB
→
Reason Ollama (local) Generate response + confidence score
→
Gate ChatDesk logic Score < 0.7 → Slack escalation
→
Billing Stripe / Paystack Metered usage tracked
◎
Multi-Modal Dashboard
Unified control surface for the full stack

The self-hosted dashboard at http://localhost:4000 is the central visibility layer — it aggregates real-time status from all 14 tools, agent session activity, board task states, and product health metrics. This dashboard (the one you're reading now) is the external, shareable version of that internal view.

Board View
Live task states from Claw-Kanban · agent assignment visibility · blockers flagged in real time
Agent Outputs
Recent completions from Claude · Gemini · Codex · SharedContext last-write timestamps
Simulations
AgentArena challenge results · AgentBenchmark leaderboard scores · active simulation status