I taught 80+ students to run their own AI agent, with the Wi-Fi off

Last week I stood in front of more than 80 students at Assumption College of Davao and did something that still feels a little bit like magic: I had an AI agent do real work on a laptop, with the Wi-Fi switched off.

This was Campus Claw, the first campus session run by my local Davao DeFi Community (DDC). I came in as the technical half of the day. Uri Buenconsejo walked everyone through the fundamentals of AI and agents, and I picked it up from there: what an agent actually is, what you can build with one, and then a live demo on my own machine.

No slides can beat the moment a room realises the "AI" is running locally, and that when I pull the network cable, it keeps going.

The one idea I wanted them to leave with

Most people meet AI as a chatbot: you ask, it answers, and you go do the work. An agent flips that. You give it a goal, and it plans, calls tools, and reports back. It can act.

The moment software can act, the whole game changes. A chatbot that is wrong wastes your time. An agent that is wrong does something. It sends the email. It spends the money. It deletes the file.

So the real question is not "which model is smartest." It is "what is this thing allowed to touch, who holds the keys, and what happens when it is wrong." Those are infrastructure questions, and a better model does not answer a single one of them. That is the whole talk in one line.

What IronClaw is

The tool I demoed is IronClaw, an open-source Agent OS built by NEAR AI. It is written in Rust, ships as one binary, and runs on your own machine. The tagline on their site sums it up: do what you do best, IronClaw will do the rest.

The part I care about most: it is local-first and privacy-first. You can run the model locally through Ollama with no API key, and your secrets never have to touch a cloud model. An agent in IronClaw is really three things:

a loop that thinks, acts, observes, and repeats
tools, which are the things it is allowed to call
memory, which is what survives after the chat window closes

And here is the teaching point students remembered: tools are permissions, not features. A tool has to declare what it can reach. Network calls only go to an allowlist. Credentials are injected at the boundary, so the tool code never even sees the key. If you have granted an Android app access to your camera, you already understand the model.

One thing I made a point of saying out loud, because a technical audience will suspect it otherwise: IronClaw is not a crypto product. I grepped every crate in the repo for wallet and chain dependencies and found zero. NEAR AI builds it, and NEAR is one of many optional model providers, but the runtime itself is just an open-source agent runtime.

The honest part

I am not going to sell you something I would not flag myself. IronClaw is early and it is rough. As I write this it has around 1,340 open issues on GitHub. I found three real bugs in a single day of using it.

But that is only half the picture. It also has 12,600 stars in about seven months, and commits land daily. It is one of the fastest-moving agent projects out there. For a student, that combination is a gift: the barrier to your first open-source contribution is on the floor. "The docs told me to run a command that does not exist" is a perfectly good first issue. That is a genuinely useful thing to tell a room full of people who think their first contribution has to be brilliant.

What you can actually build

Everything I showed runs free, on a normal laptop:

a morning briefing that runs before you wake up
a campus Telegram bot with your student handbook in its memory
automated code review on every pull request

The software is free and open source. The model is free if you run it locally. The only real floor is RAM: a model good enough for reliable tool calling wants roughly 24 GB, and most student laptops will not have that. The fix is simple, and I said so honestly on the day: use a cheap hosted model for the thinking, and keep everything else, the agent, the memory, the tool sandbox, on your own machine. That is still the whole point.

See it run

I recorded a short first look at IronClaw doing real work. If you have never watched an agent use a tool and then get told "no" by a file it is not allowed to touch, start there.

Thank you

This happened because a community made it happen. Thank you to Assumption College of Davao, the ACD League of Information Technology Students, NEAR and NEAR Legion, and my own Davao DeFi Community for putting it together. First campus down. More to come.

Run it yourself

I turned the whole session into a hands-on workshop: the slides, the setup, and a step-by-step walkthrough where every claim is sourced so you can check my work. It is free and there is no signup.

The workshop: johnforfar.com/workshops/ironclaw
The source: github.com/nearai/ironclaw

Install it tonight, ask it what tools it has, and build one thing that runs without you. That is the whole journey, and it starts on your own machine.