Executive summary
Local-first orchestration in production
6Fleet nodes
17+Models on-prem
42Skills loaded
25OpenClaw agents
100%Local inference
<10sAvg deck build
The problem we solve
Why enterprises need a local orchestrator
- Cloud LLM tokens scale linearly with traffic — operating cost grows unchecked.
- Multi-AI tools live in silos — no shared plan, memory, or accountability.
- Hallucinated "success" wastes engineering hours on returns and rework.
- Iron-law compliance needs a runtime enforcer, not a wiki page.
- Sensitive code and customer data should never leave the LAN.
What Joe does
An orchestrator built on five pillars
- Local-first inference — Ollama across RTX 5090 / Arc A770 / NVIDIA fleet.
- CEO-as-dispatcher — OpenClaw CEO routes work to the right sub-agent.
- Honest pipeline — Plan → Probe → Install → Run → VERIFY artifact ≥ 1 KB.
- Self-evolving — re-reads
/skills/joe/lessons/ before every run. - Auto-installing — missing OpenClaw agents detected and provisioned on demand.
How it works
From request to verified deliverable
One command in. Verified artifact out. No silent failures.
Fleet topology
6 nodes, hub-and-spoke, GPU-accelerated
- 1.125 — control plane · Copilot CLI · Joe
- 1.98 — RTX 5090 · nemotron · qwen3.5 · deepseek-r1
- 1.80 — Arc A770 · llama3.3 · bge-m3 · phi3
- 1.63 — NVIDIA worker
- 1.210 — worker node
- 1.10 / 1.118 — 3dd · asusdesk (LAN + Wi-Fi)
Orchestrate pipeline
Five gates between request and report
- Plan — CEO emits JSON plan
- Probe — pkgs · bins · agents
- Install — auto-yes optional
- Run — bash / python / agent
- Verify — file ≥ 1 KB on disk
- Report — CEO ratifies output
Models inventory
17+ open-weight models, deployed on-prem
| Model | Size | Role | Primary node |
|---|
| nemotron-3-super | 86 GB | Premium reasoning | 1.125 |
| llama3.3:latest | 42 GB | Long-context chat | 1.125 · 1.80 |
| nemotron-3-nano | 24 GB | Fast reasoning | 1.125 |
| joe:v1.3 | 19 GB | Joe planner | 1.125 |
| deepseek-r1:32b | 19 GB | Code reasoning | 1.125 · 1.98 |
| qwen2.5-coder:32b | 19 GB | Code generation | 1.125 |
| gpt-oss:20b | 13 GB | CEO planner | 1.125 · 1.98 |
| qwen2.5-coder:14b | 9.0 GB | Code generation | 1.125 · 1.98 |
| llama3.2-vision | 7.8 GB | Vision · multimodal | 1.125 |
| gemma2:9b | 5.4 GB | Lightweight chat | 1.80 |
| llama3.1:8b | 4.9 GB | General chat | 1.98 · 1.80 |
| qwen2.5:7b | 4.7 GB | General chat | 1.98 · 1.80 |
| gemma3:4b | 3.3 GB | Edge inference | 1.125 |
| phi3:3.8b | 2.2 GB | Edge inference | 1.80 |
| nomic-embed-text | 274 MB | Text embeddings | all nodes |
Skills catalog
42 reusable disciplines, auto-loaded before every run
Superpowers (14)
- brainstorming
- test-driven-development
- systematic-debugging
- verification-before-completion
- subagent-driven-development
- dispatching-parallel-agents
- session-recovery
- writing-plans · writing-skills
- executing-plans
- code-review (req · recv)
- self-improving-agent
- openclaw-daily-mission
- loop-self-healing
- using-superpowers
AirCore (14)
- iron-law-6 · real-browser-verify
- iron-law-10 · parallel-by-default
- iron-law-12 · ten-minute-cap
- iron-law-21 · aider-dual-track
- iron-law-29 · cpu-headroom
- iron-law-32 · real-debug-100
- iron-law-37 · frontend-debug
- ccie-enterprise-fabric
- ccie-networking-ops
- ccie-sp-and-peering
- ccie-wireless-roaming
- device-adapter-pattern
- i18n-parity-discipline
- xadmin-endpoint-cloaking
Joe + DevOps
- joe / SKILL.md (SOP)
- joe / lessons (auto-evolve)
- joe / templates (builtin)
- design-consultation · review
- qa · qa-only
- cso (security audit)
- ship · land-and-deploy
- benchmark · benchmark-models
- canary (post-deploy)
- office-hours · retro
- investigate · scrape
- iron_rules (#1-#40)
OpenClaw agent roster
25 specialised agents, grouped by function
| Group | Agents |
|---|
| C-Suite | ceo_agent · cfo_agent · coo_agent · cto_agent · cro_agent |
| Engineering | engineer_agent · cto_agent · rd-db-integrity · rnd_manager_agent |
| QA | aircore-qa · ceo-aircore-summary |
| PM / Ops | project_manager_agent · task_coordinator_agent · document_agent · meeting_agent |
| Domain — Isaac/USD | isaac-sim-expert · isaac_sim_engineer · isaac_sim_project_generator · isaac_lab_research · openusd_architect |
| Knowledge | knowledge_agent |
| Sales / CR | sales_manager_agent · client_relation_agent |
| Default / shim | main · acp-defaultagent |
Business value
What local-first orchestration delivers
- Zero cloud-token spend on inference — fixed CAPEX replaces variable OPEX.
- Data sovereignty — code · models · prompts never leave the LAN.
- Verified artifacts — no silent failures, no rework loops.
- Self-evolving — every task ends with a lesson; Joe gets smarter for free.
- Composable — Copilot CLI plans, Joe dispatches, OpenClaw executes, CEO ratifies.
- Fast time-to-value — a polished deck in under 10 seconds (this one).
Standards & governance
Iron Laws #1–#40 enforced at runtime
- #5 PostgreSQL only — no SQLite fallback
- #6 Absolute visual verification — real browser, multi-viewport
- #10 Parallel by default — multi-agent, multi-node
- #12 Ten-minute cap — no stuck tasks
- #32 Real-debug 100% — fix at the root, not the symptom
- #37 Frontend debug & evolve — chromium / mobile / iphone
- #40 Working disciplines — TDD · brainstorm gate · verify-before-done
- Source of truth:
/skills/iron_rules/iron_rules.md (878 lines)
Usage
Copy-paste ready
# Dispatch any task — Copilot CLI calls Joe
joe --orchestrate --auto-install "<task>"
# Builtin deck task (this very deck)
joe --orchestrate "generate intro pptx + html"
# Refresh fleet topology
joe --refresh-topology
# Autorun AI-suggested shell code
joe --autorun "build me a CSV of fleet RAM"
# One-shot Copilot query
joe "<question>"
# Artifacts published at:
# http://192.168.1.125/files/
Roadmap & closing
Where Joe is heading next
- Q3 — More builtin templates (report, dashboard, runbook, RCA)
- Q3 — Agent autoscale across fleet based on VRAM headroom
- Q4 — CEO ratify quorum — multi-agent agreement before sign-off
- Q4 — Self-hosted RAG over
/skills/ + project codebases - Always — Local-first · Verified · Honest
