An AI agent on a Raspberry Pi: why 50 free requests a day get you nowhere

For a while I had n8n on the Raspberry Pi, an automation tool with little boxes you wire together (“when this happens, do that”). Then I went for the trendy option: replacing it with an AI agent that would do things “intelligently”.

I installed an open-source agent, connected it to language models through a service that gives access to many models on a free plan, and ran it for a few weeks. In the end I uninstalled it. Here’s why, because I think it’s a useful lesson before you dive in.

What it promised

An AI agent isn’t just a chat: it can use tools. It reads files, runs commands, looks things up, schedules tasks… You tell it what you want and it works out the steps.

The idea was to keep it always on the Pi and ask it things from my phone: “check whether last night’s backup went well”, “summarise this week’s workouts”…

The wall: 50 requests a day

The model service had a free plan of 50 requests a day (rising to about 1,000 if you added a small amount of credit). It sounded like a lot. It isn’t.

The reason is how an agent works. A task that feels like “one question” to you is, internally, many calls to the model:

1. The model reads your request and decides which tool to use. One request.
2. The tool runs and the model reads the result. Another.
3. It decides it needs another tool. Another.
4. …and so on until it gives the final answer.

A simple task can chain ten or twenty requests. With 50 a day, the agent ran out of fuel by mid-morning.

The local fallback that didn’t work

I thought: “when the free requests run out, use a local model”. I have a laptop with a 6 GB NVIDIA GPU, so I set up an open model there.

As a local chat it was fine, but it was no use for the agent. The agent required a huge context (64,000 tokens), and with 6 GB of video memory only a small model fit, which made up the format of tool calls. For an agent, that’s like a plumber who can’t use a wrench. I cover all the numbers in another article.

What I did like

It wasn’t all bad:

• With requests available, it genuinely solved things: it read files on the Pi, summarised logs, answered with context.
• It had memory between conversations, something a normal chat doesn’t.
• It forced me to understand how agents work under the hood, which is exactly what I need for my studies.

Why I uninstalled it

In the end I did the maths:

• What I really needed were repetitive, predictable tasks: check a backup, send a weekly summary, alert me if something goes down.
• For that, a script or monitoring with alerts does exactly the same thing every time, uses no requests and makes nothing up.
• For things that do need intelligence (researching, writing, building new stuff), I use an AI assistant on my computer, when I need it, on a plan that actually covers it.

So I ended up with: a weekly script for the workout summary, Uptime Kuma with Telegram alerts for outages, and the agent gone. And the Pi with 1.7 GB fewer files.

Lessons

1. An agent burns requests like a car burns fuel in city traffic. Check the plan before you fall in love.
2. If the task is always the same, you don’t need AI. A fixed workflow (n8n, a script, cron) is cheaper and more reliable.
3. AI shines at the unpredictable: researching, writing, debugging something new.
4. A small local model doesn’t replace a big one for agents. Calling tools correctly is the hardest part, and it’s the first thing to break.
5. Uninstalling is also a technical decision. Trying something, measuring it and removing it is learning, not failing.