Last week I set a friend up with a coding agent. Linux Mint, one install command, one API key. It took about ten minutes, and most of that was the coffee.
He is doing a whole cyberpunk setup on his laptop now. When he hits a Linux problem, he pastes it into the agent and works through the fix. That is the part worth talking about: how little you actually need to get started.
You Don’t Need a Fancy Setup
People see elaborate AI setups and assume they need one too. The right subscription, a pile of tools, a weekend spent configuring everything. It can feel like there is a whole level of preparation before you get to do anything useful.
My friend’s reaction when we got it running was, “It’s that easy?”
Yeah. It can be.
You don’t need my setup. You don’t need to understand every model or build your own tools. Start with a computer you already have, install the agent, connect an account, and ask it to help with something small. You can figure out what else you need when you actually need it.
Three Pieces and You’re Started
Here is what we used:
- Linux Mint. The desktop he already had. No need to change everything before getting started.
- OpenCode. The tool you talk to when you want help with code, files, or figuring out what your computer is doing.
- OpenRouter. The account that connects it to an AI model. Add some credit, connect your key, and pick a model.
That was enough to start. He could ask questions about his machine, work through problems, and make things.
This is the same impulse behind my “just use MCP” post: check whether someone already made the thing you need. Getting started does not have to become a project of its own.
The Setup
Create an OpenRouter account, add some credit, and create an API key in its key dashboard. Then install OpenCode with the command from its official installation guide:
curl -fsSL https://opencode.ai/install | bashOpen a new terminal if opencode is not found after installation. Then log in, select OpenRouter, and paste your key at the prompt:
opencode auth loginThe CLI documentation describes this login flow. The key is saved locally so you do not need to paste it every session.
Move into an existing project directory, replacing the path below with yours, and start the agent:
cd /path/to/your-project
opencodeInside OpenCode, type /models and choose an OpenRouter model suited to coding and tool use. Its OpenRouter guide also documents /connect as an alternative way to enter your key. Many models are already listed; others may need a config entry.
Start with something small: "Explain this project and tell me how to run it. Don't change any files yet." A successful response checks that your key and selected model work before you ask for edits.
What You Give Up
Use Claude Code if it fits how you work. I would not pretend every agent has the same integrations, tool behavior, or results just because you can point it at a capable model. Those differences can matter.
And API billing is not automatically cheaper. Long sessions and large contexts can cost real money. OpenRouter's pricing page lists model access and platform fees; compare your actual usage with whatever subscription you are considering.
It is also not unlimited. Credit limits, free-model caps, and upstream rate limits still apply. Set a spending limit on your key and check the limits documentation if requests start failing.
The win for me is choice. When a model I want becomes available through the provider and works with the agent, I can try it without moving my whole workflow.
The Part That Matters
My friend has a Linux machine he wants to make his own. The agent can help inspect logs, troubleshoot Wi-Fi, write configuration files, and automate the annoying parts. That makes the terminal useful to him before he knows every command.
Access still depends on your user account and the agent's permission settings. For system changes, ask it to explain the commands and how to undo them before running them. Installing an agent does not make every proposed fix a good one.
I wrote about the same relief in AI Coding Tools Are Better at Nix Than Me. You can know what you want your machine to do and get help expressing it. You do not need Nix for this setup, either. Mint is enough.
Once you start keeping project context, my Stop Using CLAUDE.md post gets into where I put the details I want an agent to see.
There is a business version of this idea: hand a client a preconfigured agent connected to the systems they need, and they never think about the harness at all. That is the next post: Just-in-Time Agents: The Install Is the Onboarding.
This one is the floor. A Linux desktop, an installed agent, and a working API key are enough to start doing useful work.
If you just want to level up, download the boring stack. Everything else is gravy.
Header photo by Joshua Reddekopp on Unsplash.
Content on this blog was created using human and AI-assisted workflows described in my standards and workflow posts. Original ideas and editorial decisions by Justin Quaintance.