local-ai-stack

Personal AI agent infrastructure: OpenClaw Gateway + Claude Opus + Telegram Bot + private RAG (nomic-embed) + n8n. Fully operational multi-channel agent on Windows 11 — no CUDA required.

---

Stack · Metrics · Status

Model (cloud)	Claude Opus (anthropic/claude-opus-4-7) via OpenClaw
Model (local chat)	google/gemma-3-1b via LM Studio
Embeddings	nomic-embed-text-v1.5 via LM Studio
Local inference	Ollama gemma3:1b · localhost:11434
Channels	WebChat (localhost:18789) + Telegram Bot
Automation	n8n · localhost:5678
Latency (local)	<2s response · no GPU
Status	✅ Production — all services running

---

Architecture

┌─────────────────────────────────────────────────────┐
│                   CHANNELS                          │
│   WebChat (18789)        Telegram @luisAI_bot       │
└──────────────┬───────────────────┬──────────────────┘
               │                   │
               ▼                   ▼
┌─────────────────────────────────────────────────────┐
│            OpenClaw Gateway (localhost:18789)        │
│   Auth: Claude CLI · Model: claude-opus-4-7         │
│   Memory: persistent sessions · Skills: nano-pdf    │
└──────────────────────────┬──────────────────────────┘
                           │
          ┌────────────────┼────────────────┐
          ▼                ▼                ▼
┌──────────────┐  ┌──────────────┐  ┌──────────────┐
│  Anthropic   │  │  LM Studio   │  │    Ollama    │
│  Claude API  │  │ gemma-3-1b   │  │  gemma3:1b   │
│  (cloud)     │  │ nomic-embed  │  │  :11434      │
└──────────────┘  └──────────────┘  └──────────────┘
                           │
                           ▼
              ┌─────────────────────┐
              │    n8n (5678)       │
              │  Automation flows   │
              │  Claude API + hooks │
              └─────────────────────┘

---

Features

• Multi-channel agent — responds on WebChat and Telegram simultaneously
• Hybrid routing — cloud (Claude Opus) + local (Gemma via Ollama/LM Studio)
• Private RAG — nomic-embed-text-v1.5 over local .md documents, zero data leakage
• Persistent memory — OpenClaw sessions preserve context across conversations
• Automation ready — n8n connected for workflow triggers and notifications
• No CUDA — runs fully on CPU (Intel i5, 12GB RAM)
• Windows 11 native — no WSL required for core services

---

Tech Stack

Technology	Version	Purpose
OpenClaw	v2026.5.3-1	AI agent gateway, multi-channel
Claude Opus API	anthropic/claude-opus-4-7	Primary LLM (cloud)
Telegram Bot API	—	Messaging channel
Ollama	v0.22.1	Local LLM inference
LM Studio	v0.4.12	Local chat + embeddings server
nomic-embed-text	v1.5	Private RAG embeddings
gemma3:1b	—	Local chat model (815 MB)
n8n	latest	Automation & workflow engine
Claude Code	—	CLI auth for OpenClaw

---

Quick Start

# 1. Clone
git clone https://github.com/lcarrenoy/local-ai-stack.git
cd local-ai-stack

# 2. Install OpenClaw
npm install -g openclaw

# 3. Authenticate Claude CLI
npm install -g @anthropic-ai/claude-code
claude auth login

# 4. Configure
cp .env.example .env
# Edit .env with your Telegram bot token

# 5. Run onboarding
openclaw onboard

# 6. Start gateway (as Administrator)
openclaw gateway install
openclaw gateway start

# 7. Open dashboard
start http://127.0.0.1:18789

---

Services & Ports

Service	Port	Start Command	Auto-start
OpenClaw Gateway	18789	openclaw gateway start	✅ Scheduled Task
Ollama	11434	auto	✅ always running
n8n	5678	n8n start	⚠️ manual
LM Studio	1234	open app → Start Server	⚠️ manual
MLflow	5000	mlflow ui --host 127.0.0.1 --port 5000 --workers 1	⚠️ manual

---

Project Structure

local-ai-stack/
├── config/
│   ├── openclaw.json.example     # OpenClaw gateway config
│   └── telegram_setup.md         # Bot setup guide
├── docs/
│   ├── architecture.md           # Full architecture notes
│   └── setup_windows.md          # Windows 11 setup guide
├── models/
│   └── models_info.md            # Downloaded models reference
├── n8n/
│   └── workflows/                # n8n workflow exports (.json)
├── rag/
│   └── docs/                     # Documents for RAG ingestion
├── .env.example
├── .gitignore
└── README.md

---

Key Results

Metric	Value
Response latency (local)	<2s on Intel i5 + 12GB RAM
Channels active	2 (WebChat + Telegram)
RAG documents	Private local .md files
GPU required	❌ None (CPU-only)
Models running	3 (gemma3:1b + gemma-3-1b + nomic-embed)

---

Roadmap

• Add WhatsApp channel via n8n + Twilio
• RAG over financial documents (PDF ingestion)
• LLMOps evaluation with DeepEval + LangFuse
• Connect to AI Core System (LangGraph multi-agent)
• Add more Ollama models (Llama3, Mistral)
• Skills: obsidian, session-logs via WSL

---

Requirements

• Windows 11 (or WSL2 for Linux-based skills)
• Node.js 22+
• Anthropic account (Claude Pro or API key)
• Telegram account (for bot channel)
• 12GB RAM minimum (16GB recommended)
• No GPU required

---

Part of Luis Carreño's portfolio · AI Engineer · Financial Engineering · Score 9.8/10