The stack worth knowing — open-source, free, and changing what one person can build.
You don't need a lab or a budget to build with AI anymore. The best models run on a laptop, the best coding tools are free, and the licences let you use almost all of it commercially. This is the map I wish I'd had starting out — every tile says what it is, why it matters, and the licence, because knowing what you're allowed to do with something is half the lesson.
How to read a licence (the 60-second version)
Open source isn't one thing — the licence decides what you can build on top. Four you'll see everywhere:
Models you can download and run for free.
Llama
Meta's family — the one that started the open-weights wave. Sizes from runs-on-a-phone to runs-a-cluster. The safe default for general work.
DeepSeek (V3 · R1)
The one that shocked everyone — frontier-class reasoning at a fraction of the training cost, weights released MIT. Proof open can catch up fast.
Qwen
Alibaba's family — strong at code, maths and non-English. Most sizes are Apache-2.0, so you can ship commercially with no asterisks.
Mistral & Mixtral
French lab, famous for small models that punch far above their weight. Mixtral pioneered cheap “mixture-of-experts” for the open world.
Gemma
Google's open siblings to Gemini. Small, fast, genuinely good — built to run on your own hardware, including phones and single GPUs.
Phi
Microsoft's “small but mighty” line — trained on textbook-quality data so a tiny model reasons like a much bigger one. MIT-licensed.
Hugging Face
The GitHub of AI models — a million-plus models, datasets and demos in one place. Where every model above actually lives. Start here.
Whisper
Speech-to-text that just works, in ~100 languages, fully open. It's the engine inside my own murmur app — your voice never leaves the machine.
Get a model running on your own machine.
Ollama
One command — ollama run llama3 — and you have a local model. We use it for first-pass work before spending a single paid token.
LM Studio
A friendly desktop app to download and chat with open models — point, click, run. The gentlest on-ramp if the command line scares you.
llama.cpp
The engine under most local tools. Squeezes big models onto plain laptops — no fancy GPU required. The reason any of this works on cheap hardware.
vLLM
The serious server engine — what you reach for when you need to serve a model to a whole product fast. Powers production deployments everywhere.
Open WebUI
A polished ChatGPT-style interface that sits on top of Ollama. Self-host it and your whole team has a private chat — your data, your server.
LumiChats (offline)
Private AI on a Windows PC — no internet, no GPU, no cloud. Nine fine-tuned models, built on GPT4All. A fork I keep for fully air-gapped work.
The tools that change what one person can ship.
Claude Code
An agent that lives in your terminal, reads your whole codebase, and does the work — not autocomplete, a teammate. This entire site was built with it.
Aider
The open-source pair-programmer in your terminal. Bring your own model (even a local one), and it edits your repo and commits to git for you.
Cline
An autonomous coding agent right inside VS Code — it plans, edits files, runs commands and asks before anything risky. Fully open, model-agnostic.
Continue
Open-source Copilot you fully control — pick any model, point it at local ones, build custom assistants. The configurable, no-lock-in option.
Cursor
A whole code editor rebuilt around AI — chat with your codebase, multi-file edits, agent mode. Not open, but the free tier is where many people start.
Open Browser Use
Open alternative to letting an agent drive a real browser — click, type, fill forms, complete tasks on the live web. A fork I keep for web automation.
Frameworks that make models do things, not just talk.
Model Context Protocol
The “USB-C for AI” — one open standard so any model can plug into any tool, database or app. The most important plumbing in the field right now.
MCP Servers
The reference collection of ready-made MCP connectors — GitHub, Slack, databases, the web. Fork it, run one, and your agent gains a new sense.
LangChain / LangGraph
The most-used framework for chaining models, tools and memory into real apps. LangGraph adds proper, stateful agent workflows on top.
CrewAI
Build a “crew” of role-playing agents that collaborate on a task — researcher, writer, reviewer. The simplest way to grasp multi-agent work.
Agent Development Kit
Google's open framework for building and shipping production agents — the same one their own teams use. Clean, opinionated, well-documented.
Browser Harness
A self-healing layer that lets a model finish any task in a real browser, recovering when a page changes. One of my bets on web-acting agents.
Make your AI cheaper, faster, and able to remember.
Glia-AI
A persistent local memory layer that syncs context across Claude, ChatGPT, Cursor and more — into one local knowledge graph. The “stay synced across tools” piece.
Code Review Graph
Builds a persistent map of a codebase so the AI reads only what matters — up to 49× fewer tokens on daily tasks. Massive cost saver on big repos.
caveman
“Why use many token when few token do trick” — a Claude Code skill that cuts ~65% of tokens by talking terse. Funny name, real savings.
rtk1
A tiny CLI proxy that reduces LLM token use 60–90% on common dev commands. One Rust binary, zero dependencies. Sits quietly and saves money.
claude-context
Code-search over MCP — makes an entire codebase searchable context for any agent, so it finds the right file instead of reading everything.
everything-claude-code
A whole performance system for coding agents — skills, instincts, memory, security, research-first defaults. The playbook I mine for ideas.
Where it's going next — from robots to physical AI.
LeRobot
Hugging Face's open robotics stack — models, datasets and tutorials to teach a real robot arm new skills. The cheapest door into physical AI.
Genesis
A blazing-fast open physics engine for training robots in simulation before they ever touch the real world. Generates training worlds from a prompt.
Unitree
The company making humanoid and dog robots people can actually afford — and they open-source SDKs and training code. Where hobbyist robotics is heading.
NVIDIA Isaac & GR00T
NVIDIA's platform for robot foundation models — train, simulate, deploy. GR00T is their open base model for humanoids. The picks-and-shovels play.
Figure
One of the leaders racing to put general-purpose humanoid robots into real factories and homes. Closed, but the bellwether for how fast this is moving.
ComfyUI
A visual, node-based studio for open image and video models — drag boxes, wire them up, generate. The open creative counterweight to closed tools.
The bets I keep on my own shelf.
superpowers
An agentic skills framework and development method that actually works in practice — a structured way to give agents reusable abilities.
openhuman
A private, simple, powerful personal super-intelligence — the “your own assistant, fully yours” idea I keep circling back to.
DocuSeal
Open-source DocuSign alternative. How we get HIPAA agreements signed on our own server, no per-seat bill. Already running in production for us.
repowise
Codebase intelligence for AI teams — auto-docs, git analytics, dead-code detection and architecture notes, all over MCP. Keeps big repos legible.
Transformer Explainer
An interactive, visual way to see how an LLM actually works inside. If you only open one link to learn the fundamentals, make it this one.
github.com/jakesparow
The full shelf — 35+ forks across models, agents, security, OCR, Kubernetes and more. Browse it like a reading list of what I think matters.
Pick one. Build something tiny this week.
You don't learn this by reading the list — you learn it by running ollama run llama3 once, or letting an agent fix one bug. Start small. The rest follows.