Build once. Govern anywhere.
Built at AI Tinkerers Hackathon · Durango, Mexico 2025
ForgeOS3 is not another agent runtime. It is a reusable infrastructure layer that sits on top of existing agent frameworks and makes them safe, auditable, and production-ready.
While frameworks like LangGraph, AutoGen, and CrewAI focus on how agents think, ForgeOS3 focuses on what they're allowed to do — and ensures every decision is enforced, logged, and explainable.
[ Your Agent Runtime ]
↓
[ ForgeOS3 Infrastructure Layer ]
Policy Engine · Tool Gateway · Loop Guard
Approval Router · Sandbox · Audit Trail
↓
[ Supabase · Event Log · Approval Queue ]
| Module | Description |
|---|---|
| 🏗️ Builder Console | Create agents from reusable templates with domain profiles, tool packs and policy presets |
| 📦 Registry Manager | Centralized store of domain profiles, tool packs, policy configs and runtime adapters |
| ⚖️ Policy Engine | Evaluates every tool intent and returns allowed, blocked, or approval_required |
| 🔀 Tool Gateway | Intercept layer — every tool call passes through before execution |
| 🔁 Loop Guard | Detects runaway behavior, tracks loop risk scores, activates safe mode or kill switch |
| 🧱 Sandbox Layer | Isolated execution with timeouts, resource limits, network allowlist and secret scoping |
| 👁️ Sentinel Studio | Live dashboard of active runs, tool timelines and decision badges |
| 📋 Audit Trail | Full institutional record of runs, decisions, approvals and risk events |
| ✅ Approvals Panel | Human-in-the-loop workflow for sensitive tool calls |
| 🔌 Runtime Adapters | Connect to any agent framework — OpenClaw (live), LangGraph, AutoGen, CrewAI (planned) |
The same agent behaves differently depending on its domain profile and policy preset.
Scenario A — Health
Input: "Summarize patient intake #4821 and create a follow-up checklist"
summarize → ✅ allowed
checklist → ✅ allowed
diagnose → ❌ blocked (critical sensitivity · strict policy)
Scenario B — Gov
Input: "Analyze public request #2291 and route to the right department"
classify → ✅ allowed
route → ✅ allowed
write_external → ⏳ approval required (writes to municipal DB)
Scenario C — Marketing
Input: "Generate a campaign workflow and prepare a content draft"
summarize → ✅ allowed
draft → ✅ allowed
publish → ⏳ approval required (external channel publish)
Scenario D — Loop Guard
Same tool called repeatedly → risk score escalates → safe mode activated → kill switch
┌─────────────────────────────────────────────────────────┐
│ ForgeOS3 Web App │
│ Builder · Registry · Policy · Sentinel · Audit │
└──────────────────────────┬──────────────────────────────┘
│ HTTP
┌──────────────────────────▼──────────────────────────────┐
│ ForgeOS3 Core API │
│ Policy Engine · Tool Gateway · Loop Guard │
│ Approval Router · Sandbox Layer · Audit & Event Log │
└──────────────────────────┬──────────────────────────────┘
│
┌──────────────────────────▼──────────────────────────────┐
│ Runtime Adapter Layer │
│ OpenClaw Adapter (live) · LangGraph · AutoGen (planned) │
└──────────────────────────┬──────────────────────────────┘
│
┌──────────────────────────▼──────────────────────────────┐
│ Agent Runtime(s) │
│ OpenClaw · Custom Enterprise │
└─────────────────────────────────────────────────────────┘
forgeos3/ ← Frontend (Cristian)
├── src/
│ ├── pages/ # 13 pages — Builder, Sentinel, Audit, etc
│ ├── components/ # UI system — Badge, Button, Card, Modal...
│ ├── store/ # Zustand — auth, agents, runs
│ └── types/ # Shared TypeScript types
forgeos3-backend/ ← API + Core Engine (Diego)
├── src/
│ ├── routes/ # REST endpoints
│ ├── engine/ # Policy, Gateway, LoopGuard, Audit, Sandbox
│ └── db/ # Supabase client + seeds
forgeos3-agent/ ← OpenClaw Adapter (William)
├── src/
│ ├── adapter/ # RuntimeAdapter implementation
│ ├── scenarios/ # Health, Gov, Marketing demo scenarios
│ └── tools/ # Tool definitions per domain
- Node.js 20+
- A Supabase project
- OpenClaw installed
git clone https://github.com/diegoramirez772/forgeos3.git
cd forgeos3cd forgeos3-backend
npm install
cp .env.example .env # add your Supabase credentials
npm run dev # runs on http://localhost:3001cd forgeos3
npm install
cp .env.example .env # set VITE_API_URL=http://localhost:3001
npm run dev # runs on http://localhost:5173cd forgeos3-agent
npm install
cp .env.example .env # set FORGEOS3_API_URL=http://localhost:3001
npm run demo:health # run the Health scenario| Layer | Technology |
|---|---|
| Frontend | React 19, Vite, TypeScript, Tailwind CSS, Zustand, Framer Motion |
| Backend | Node.js, Express, TypeScript, Zod |
| Database | Supabase (PostgreSQL) |
| Agent Runtime | OpenClaw |
| Shared | TypeScript interfaces across all three projects |
| Role | Person |
|---|---|
| Backend · DB · Core Engine | Diego |
| Runtime · OpenClaw Adapter | William |
| Frontend · Dashboard · UI | Cristian |
- Frontend scaffold — all 13 pages built
- Mock data layer — full demo with realistic data
- Supabase schema + seeds
- Core API endpoints
- Policy Engine logic
- OpenClaw Adapter integration
- Full end-to-end demo (4 scenarios)
- Deploy
ForgeOS3 · AI Tinkerers Hackathon · Durango, Mexico 2025
We are not building another agent. We are building the infrastructure layer that makes agents safer, reusable, and production-ready across runtimes.