masterofsomething

In Practice Media

Tools

The actual stack. 20 tools across 5 categories. Everything here is something we run on the system this site describes. No affiliate padding — if a tool is on this page, we use it.

Local AI & Models

The runtime layer — what actually executes the model on your hardware.

Runs LLMs locally — model download, inference, simple HTTP API.

The fastest path to a private AI. One install, one command, working chat.

Self-hosted ChatGPT-style interface that pairs with Ollama.

Persistent history, model switching, file uploads. The interface you live in once the terminal stops scaling.

openwebui.com ↗

Desktop app for running local models with a polished GUI.

A more approachable on-ramp than Ollama if you prefer GUI over CLI. Model browser is solid.

lmstudio.ai ↗

All-in-one local AI workspace with documents, agents, and chat.

If you want a batteries-included desktop app rather than assembling parts. Good for non-technical household members.

anythingllm.com ↗

Memory & Knowledge

How the AI remembers things across sessions.

Markdown-first knowledge base. Notes are plain files, indexed for FTS + vector search.

The simplest possible persistent memory layer. You can read and edit your AI's memory in any text editor.

github.com/basicmachines-co/basic-memory ↗

Memory layer for AI apps — store, retrieve, and decay conversation snippets as embeddings.

Pairs with Qdrant for semantic recall. The right tool when you need "what did we discuss three weeks ago" search.

High-performance vector database. Runs in Docker, no auth needed for local use.

The vector store under our memory stack. Faster than the alternatives and zero-config for a single-user setup.

qdrant.tech ↗

Graph database with Cypher query support. Drop-in for Neo4j-style use cases.

Backs our Graphiti causal graph — entity relationships across time without a JVM.

falkordb.com ↗

Agents & Automation

Tools that turn your AI into an actor, not just a responder.

Self-hosted workflow automation — visual editor, hundreds of integrations.

The glue between your AI, your inbox, your calendar, and your home. Schedules, triggers, webhooks.

Anthropic's autonomous developer agent — drives a real terminal, edits real files.

The build agent. We use it to write code, deploy infrastructure, run system tasks. Full agentic envelope, not chat-with-snippets.

claude.com/product/claude-code ↗

Node-based interface for running image generation models locally.

When the cost of cloud image gen adds up, this is the local alternative. Steeper curve, total control.

github.com/comfyanonymous/ComfyUI ↗

Browser automation framework — Chrome, Edge, Firefox via a single API.

Our browser-agent foundation. Drives ChatGPT, Substack, GitHub, anything web — when you need an AI to act in a real browser session.

playwright.dev ↗

Infrastructure

The substrate everything runs on.

Zero-config private network across all your devices via WireGuard.

Reach your home AI from your phone, your laptop, anywhere — without opening ports or trusting a cloud relay.

tailscale.com ↗

DNS, CDN, Pages hosting, free SSL, privacy-first analytics.

masterofsomething.com runs on Cloudflare Pages. Free tier handles real traffic. The DX is unusually good.

cloudflare.com ↗

Container runtime — package and run apps in isolated environments.

Qdrant, FalkorDB, n8n, Open WebUI — all of these run as Docker containers on the same machine. Cleaner than installing each one natively.

Local-first home automation platform. Hundreds of integrations.

The substrate the AI plugs into for any physical-world action. The only mature option for a privacy-respecting smart home.

home-assistant.io ↗

Models Worth Knowing

The actual brains. Pull these from Ollama or HuggingFace.

Alibaba's open-weights model family. The 8B variant is our daily driver locally.

Best quality-per-VRAM in the 7-9B range as of 2026. Multilingual, good at reasoning, fits in 8 GB.

qwenlm.github.io ↗

Meta's flagship open model. 8B and 70B variants commonly used.

Solid alternative to Qwen. The 70B (quantized) is what you run if you have 12+ GB VRAM.

ai.meta.com/llama ↗

Open-weights image generation model from Black Forest Labs.

Best open image model in 2026. The schnell variant fits on 8 GB VRAM at 8-bit. Local image gen without an API bill.

blackforestlabs.ai ↗

OpenAI's open-weights speech recognition. Variants like faster-whisper run locally with low latency.

The voice-input layer. Runs on CPU well enough for a single user; on a GPU it's near real-time.

github.com/openai/whisper ↗