What to actually use
The hardest part of starting is that every option sounds equally good. This is the opinionated version: what we reach for, what it's for, and what you can safely ignore until you have a reason not to.
Models
Any of the current frontier models from Anthropic, OpenAI or Google will do everything on this site. Pick one, learn its quirks, and switch only when something forces you to — the differences that matter to a beginner are far smaller than the internet suggests.
The choice that actually matters is big model or small model. Use a small, cheap, fast model for classification, extraction and routing — the things you do thousands of times. Use a large one for judgment, writing and anything multi-step. Most projects that get expensive got that way by sending everything to the big model out of habit.
Run something locally (Ollama and a small open model) when the data can't leave the building, when you need it free at high volume, or when you're on hardware without reliable internet. Otherwise a hosted API is cheaper than your time.
Frameworks
Start with no framework. An HTTP call to a model API in fifty lines of Python teaches you more than any abstraction, and a surprising number of production agents never need more than that.
When you do need structure — branching, retries, multiple agents, resumable state — LangGraph is the most solid choice today; CrewAI is friendlier if your problem really is several agents with distinct roles. MCP is worth learning as soon as you want an agent to use tools, because it's becoming the common way to expose them and it saves rewriting integrations per model.
For visual and no-code building, n8n is the one to learn: self-hostable, handles real branching, and the concepts transfer to code later. Zapier and Make are fine for simple linear automations.
Boards
| Board | Roughly | Use it when |
|---|---|---|
| ESP32 | $5–12 | Sensors, small always-on devices, battery power. No operating system — it just runs your code. Start here for anything that reads a sensor and sends a message. |
| Raspberry Pi Zero 2 W | $15–25 | You need Linux, a camera or Python libraries, but the job is small. The sweet spot for a always-on camera or a tiny server. |
| Raspberry Pi 5 | $60–100 | Real computing — running models locally, handling audio and video, or acting as the brain for something bigger. |
Prices vary by region and retailer and have moved around a lot; treat these as rough bands, not quotes.
Hosting and running things
Anything that needs to run on a schedule can start as a GitHub Action on a cron — free, versioned, no server to maintain. Move to a small VPS when you need something always-on or holding state. A Netlify or Vercel function is the easiest way to put a small API behind a website.
For storage, SQLite is enough for far longer than you'd think, including as a vector store for a few thousand documents. Reach for Postgres with pgvector when you outgrow it, not before.
What you can ignore for now
Fine-tuning — almost every problem a beginner thinks needs it is solved by a better prompt or retrieval. Vector database services, until SQLite genuinely isn't coping. Multi-agent architectures, until one agent has demonstrably failed at the job. Agent observability platforms, until you have users. Every one of these is a real tool for a real problem you don't have yet, and adopting them early is the main reason projects stall before shipping.