sns-network-solutions/ai-core

AI Core — SnS Network Solutions

The holding company's internal AI team. A strict hierarchy: one leader agent (Chief) is the single point of contact and routes work to a lean set of specialists. Nothing runs unless it's necessary to the business.

Goal: all services eventually run on AWS (EC2 + Bedrock). For now it runs on the home server ws in a VM. The stack is containerized so it moves between the two unchanged — the VM is disposable; this git folder is the source of truth.

Chain of command

        SAM  (owner / CEO — final authority, approval gates)
         │   you talk to ONE agent
        CHIEF  (team leader — routes, decides, aggregates, enforces gates)
         │   specialists report UP only, never to each other
   ┌─────┼─────┐
 SCOUT  SCRIBE  (LEDGER — add when billing starts)
research  docs

• You only ever message Chief. Specialists have no direct channel to you.
• Specialists only talk to Chief — no lateral messaging (enforced by config).
• Chief cannot send external comms, move money, or make legal promises — those escalate back to you. Chief runs the team; you stay CEO.

Roster (lean — only what the business needs now)

Agent	Role	Status
Chief	Team leader / orchestrator. Single point of contact, routing, aggregation, approval gates.	Active
Scout	Research & competitive intel. Read-only, informational.	Active
Scribe	Documents — proposals, SOWs, NDAs, SOPs — from templates, held for review.	Active
Ledger	Bookkeeping/billing prep. Generates, never sends or pays.	Add when invoicing starts

Add specialists under Chief as the business creates work for them. The command structure never changes — that's the point.

Layout

ai-core/
├── docker-compose.yml     OpenClaw gateway (image ghcr.io/openclaw/openclaw, :18789)
├── .env.example           config template — copy to .env, fill in, never commit .env
├── .gitignore
├── agents/                canonical agent definitions (version-controlled)
│   ├── chief/             team leader
│   ├── scout/             research specialist
│   └── scribe/            document & visual specialist
│       each agent folder holds its OpenClaw workspace:
│         SOUL · IDENTITY · USER · AGENTS · TOOLS · HEARTBEAT · MEMORY
│         + skills/   knowledge/   projects/
├── knowledge/             shared knowledge base — sns.md (brand foundation) + research/reference docs
└── state/                 runtime state + sessions (gitignored, created on first run)

Run it (on the ws VM)

Prereqs: a Linux VM with Docker + Docker Compose, and ≥4GB RAM (OpenClaw OOMs under 2GB).

cp .env.example .env        # then fill in your OpenAI API key + model refs
docker compose up -d
docker compose logs -f gateway

The gateway binds to 127.0.0.1:18789 only — it is not exposed to the internet. Reach the web UI over Tailscale or an SSH tunnel:

ssh -L 18789:127.0.0.1:18789 <user>@ws   # then open http://localhost:18789

Register the agents (first run) — the canonical SOUL.md files live in agents/; register each with the gateway so it picks them up:

docker compose exec gateway openclaw agents add chief
docker compose exec gateway openclaw agents add scout
docker compose exec gateway openclaw agents add scribe
docker compose exec gateway openclaw agents list --bindings

ponytail: exact agents add workflow + workspace path (~/.openclaw/agents/<id>/workspace/SOUL.md) should be confirmed against https://docs.openclaw.ai on first run — ceiling: this is a one-time manual registration step, not yet automated. Upgrade path: a bootstrap script that syncs agents/*/SOUL.md into the state volume on up.

LLM backend — OpenAI / ChatGPT (current)

OpenClaw talks to OpenAI via an API key.

• On ws now: put your OPENAI_API_KEY in .env (see .env.example) and set the model refs. Route routine work to a small/fast model and reasoning to a strong one to control cost.
• The key is the only credential needed — no cloud account required to run on ws.

Move to AWS + Bedrock (future goal)

The end goal is to run this same stack on AWS with Amazon Bedrock. When we migrate, run the same docker compose up on an EC2 instance (t3.large, 8GB): drop the OPENAI_* keys, attach an IAM instance role so OpenClaw uses Bedrock via the AWS credential chain, put a Bedrock VPC endpoint in front, and reach the gateway over Tailscale/SSM. The agents, skills, and compose file are identical — only the provider config changes. That backend portability is the whole point of this design.