AI Company Brains and Personal Agents

AI Company Brains and Personal Agents

AI Tinkerers OC — June 2, 2026 | 10-min demo

What This Is

An operating system for AI agents running entirely on local hardware — zero cloud LLM spend. 196 microservices, 43 agents, 147 MCP tools.

Live right now:

• portal.aitherium.com — agent fleet hub
• chelle.aitherium.com — photography business AI
• garg.aitherium.com — government consulting AI
• irc.aitherium.com — live chat (join #tinkerers!)
• Discord: Vibeful — AI dev community

RTX 5090 + DGX Spark under my desk. Demoing through a Cloudflare tunnel back to my house.

---

The Core Insight: Two-Tier Routing

One model handles everything. Reasoning is a tool, not a tier.

User Request
    |
    v
[Nemotron-Orchestrator-8B] ← always resident, 14GB with TQ4
    |
    |--- 90% of requests → direct response
    |
    |--- "I need to think harder" → calls deep_reasoning tool
    |                                    |
    |                                    v
    |                             [DeepSeek-R1:14b]
    |                                    |
    |                                    v
    |--- ← reasoning result ← ─────────┘
    |
    v
Response

Why it works:

• 8B model handles tool calling, routing, orchestration, and most responses
• Reasoning model only fires when the orchestrator decides it needs deep thinking
• Both run on one RTX 5090 via vLLM continuous batching
• 4-8 agents can work simultaneously on one GPU

AitherKVCache

4-bit KV cache compression → 3.8x memory savings → 6x concurrent requests at 40K context.

pip install aither-kvcache
# Drop into any vLLM instance:
# --kv-cache-dtype tq-t4nc

---

Claude Code as Command & Control

Claude Code is my brain. AitherOS is my local AI infrastructure.

147 MCP tools — all routed to local compute:

• Code search (CodeGraph + Repowise)
• Agent dispatch (forge_subagent)
• Memory graph (remember, recall, query_memory)
• Git, filesystem, terminal
• LLM inference (local vLLM)

Claude Code → MCP → AitherNode → local services → vLLM → response

The key: Cloud intelligence for planning, local hardware for execution.

---

Managing the Infrastructure

AitherZero: 170+ PowerShell Scripts

Numbered automation scripts organize the entire lifecycle:

00-prerequisites/    # Environment setup
10-initialization/   # First-time config
20-networking/       # Tunnel, DNS, certificates
30-services/         # Docker compose orchestration
40-deployment/       # Build, push, deploy
50-maintenance/      # Health checks, cleanup
60-monitoring/       # Grafana, alerts
70-security/         # Vault rotation, audit
80-testing/          # Pester, pytest, integration
90-cleanup/          # Teardown, reset

Docker Compose Profiles

Graduated startup — don't boot 90 containers for chat:

# Raw LLM chat (~20 containers)
docker compose --profile chat-minimal up -d

# Full personality + memory (~29 containers)
docker compose --profile chat-full up -d

# + Agent tool use (~31 containers)
docker compose --profile chat-agents up -d

API Contracts as Boundaries

Every service follows the same pattern:

import services._bootstrap
from lib.core.AitherIntegration import AitherService, get_port
from lib.core.AitherEvents import setup_lifecycle

aither = AitherService("MyService", port=get_port("MyService"))
app = aither.app
setup_lifecycle(app, "MyService", port=PORT, register_identity=True)

config/services.yaml is the single source of truth for all 196 services.

---

What You Can Take Home

Install

pip install aithershell aither-adk aither-kvcache

aithershell — Standalone CLI

Works with Ollama, vLLM, or OpenAI directly (no AitherOS required):

aither --init                    # Configure your LLM backend
aither "build me an agent"       # Instant chat

# Or set backend explicitly:
export AITHER_LLM_BACKEND=ollama
export AITHER_LLM_URL=http://localhost:11434
aither "hello"

aither-adk — Agent Development Kit

aither init myagent              # Scaffold agent project
cd myagent
aither run                       # Start agent server

# Auto-detects: Ollama → vLLM → OpenAI
# Generated config.yaml defaults to llm_backend: auto

aither-kvcache — KV Cache Compression

Drop into any vLLM instance for 3.8x VRAM savings:

pip install aither-kvcache
# Add to vLLM args: --kv-cache-dtype tq-t4nc

---

Hardware Scaling

---

Pitfalls

• Don't use one giant model — small orchestrator + reasoning-as-tool scales better
• Don't skip KV cache compression — it's free 3.8x VRAM savings
• MCP tools are the composability layer — build there, not from scratch
• Docker compose profiles — don't boot everything for development
• API contracts — services.yaml + AitherService pattern keeps 196 services manageable

---

Links

• GitHub: aither-adk | aithershell | aither-kvcache | AitherZero
• Chat: irc.aitherium.com (join #tinkerers)
• Discord: discord.gg/PpEFUVHv (Vibeful — AI dev community)
• Portal: portal.aitherium.com
• PyPI: pip install aithershell aither-adk aither-kvcache