--- date: '2026-10-06' description: 'What I use every day to work and learn: an agent that lives on a Raspberry Pi (before that, on an old phone with Termux) and a wiki of text files. No hype and with the cost receipts.' lang: en readingMinutes: 8 tags: - ai - hermes - learning title: MacGyver on a Raspberry Pi url: https://foron.dev/en/blog/macgyver-on-a-raspberry-pi/ --- Many of the things that are worth it come from curiosity. Others come from need. What I am telling here sits in the middle of the two. Among so much slop on the internet, I realised I could not keep up with every information channel. It was adapt or die, in the purest Darwinian style. Madafaka. Since the start of August I have been putting everything I read and everything I decide into a folder of text files. Today it holds 471 atomic notes, 333 sources, 79 syntheses and 7 exchanges. An hour ago it held 470 and one fewer, because the system keeps working while I do other things. None of that is a product. It is plumbing and a habit. Before the Raspberry Pi, all of this lived on an old phone running Termux, which is still as good a way as any of having a server. ## What I have set up The agent is Hermes. It runs as a user service on a Raspberry Pi I keep at home, no root and no containers, and I talk to it over Telegram from my phone [^1]. That is the whole interface. There is no dashboard and no vector database. The wiki is a folder of markdown files. I open it in Obsidian on the laptop and a git repository carries it, so every change has an author and a date. Every file has a type. A loose idea goes into an atom, and whatever comes from outside goes into a source. When several ideas come together into a synthesis I write a molecule. The exchanges keep the conversations that ended in a criterion. The agent has 152 procedures saved as skills. Some write notes, others deploy services. When I say it works on its own, this is what I mean: it knows how it is done and it does it with the hands I gave it. And the numbers I care about most: 504 commits since 4 August, 61 of them so far in October. I have written almost none of them.
Commits per week in the wiki repository and the monthly cost of the deployment, priced at list rates
Commits per week in the wiki repository and what the deployment costs, priced at list rates: August, September and the first six days of October. The data comes from git itself and from the agent's own usage log.
## The loop The part that is not mine came from a video about agents and agent teams [^2]. I kept nothing from the diagram. What I copied was where the approval point sits. It works like this. A source arrives. Yesterday it was a Twitter thread with ten memory projects for agents; other times it is an article somebody sends me. The agent reads it, looks at what is already in the folder, decides where it fits, writes a draft and hands it back. I read it. Sometimes I tell it no. Other times I tell it what to change. And not every no is about judgement: the agent is a non-deterministic program, some days it answers me in Chinese and I have to steer it back, and other times it suggests something plainly false, or dangerous to run. If I say nothing, it gets filed and connected to what was already there. Curation is that conversation, not a process that runs on its own and reports back at the end with a summary of what it did. The thread about the ten projects came out of there turned into something else: the repositories verified one by one, the axes that really separate them and a criterion written in seven points. That is not a summary of the thread. It is what was left after fighting it. And this moves forward when I am not at the desk. Once my master's was done, I kept adding features to Norn from the beach: the instructions and the approvals came out of the phone, and the commits and the notes came out of home. The approval point travels with me. You do not need a project either: I can even generate spreadsheets, if I ever fancied it. ## What fills up without me There are eight scheduled tasks. At seven in the morning the agent goes through the security feeds, filters out what is already covered and writes the summary of the day. Every six hours it reviews the links in the new notes and warns about the ones pointing at nothing. Another one looks for orphan notes twice a day, and one syncs the folder every two hours, in case I write from the phone. At nine it leaves me a digest of what moved in the wiki, and on the first of each month, a summary of the month. The last two do not touch the wiki: one reminds me of the deadlines I have on top of me and the other warns me if the board is running out of disk. I am writing this so it is clear it is not magic: it is a few lines of configuration and a clock. And the daily summary is not knowledge, it is a signal that something happened: the atoms pile up with nobody looking at them, and making them count is my work, reading them, throwing half away and keeping what changes something. If that is not done, the folder grows and serves no purpose, which is the most comfortable way of failing at all this. ## Where judgement is formed A summary is consumed in two minutes and leaves nothing behind. Judgement is formed by arguing it, so the notes that serve me best are the ones that keep the path: the question, the answer and where it landed. And what this system rewards is consistency. The indexing starts gaining power when what is new collides with earlier ideas and opens paths where there were none. It is the best example I have of the concept «knowledge compounds», knowledge that composes instead of piling up. If what counts is accumulated judgement, the consequence is uncomfortable: any shortcut that saves me reading or thinking is a loss, even when it is comfortable. The agent is useful to me when it puts the material in front of me or argues an answer with me. If it hands it to me chewed, it is taking the exercise away from me. ## What it costs and what I do not buy The cost is real and it comes in three places. The first is money. Keeping this standing costs the OpenCode Go subscription, ten dollars a month, plus the electricity [^3]. I do not count the Raspberry, it was already at home. Every task the agent runs burns tokens, and some tasks are not worth what they cost: I learned to check the quota before launching anything long, because one unsupervised night of loop takes away what you would not spend in a month of work. With a better machine I would have a model running at home all day, but for now the world is not perfect. The second is maintenance. Links break, file names drift out of sync, and once the phone moved a note out of its folder into the root of the vault without saying anything. That kind of failure is invisible to any automatic detector, because the links still work. The third is the hardest to see: the feeling that something works because there are a lot of files. A pile of notes is not ordered knowledge. I lied to you. There is a fourth cost, and it appears on no invoice. This enables so much that it hooks you: the system does not get tired, you drop material on it and it cross-references it, and with everything within arm's reach everything feels possible. The ceiling is not set by the machine. In this system the limit of comprehension is the bottleneck: what I do not understand gets filed just like what I do understand, takes up the same room and is good for nothing. You have to know when to stop, and much to my regret, we are the limiting link. ## Closing This is the most talked-about thing of the year, so the value is not in being amazed. It is in looking under the interfaces that look like they came out of Iron Man and the AI-bro slop. Down there is what actually works: an agent running as a service on a board, a folder of text in git and a habit that is the hard part, deciding what goes in, approving what stays and throwing away what does not. I built this to learn faster, not to write less. And with that I am back to the usual, which is the only thing I have not managed to automate: choosing what is worth it. ## References [^1]: [Hermes Agent](https://github.com/NousResearch/hermes-agent) · Nous Research [^2]: [Subagents vs Agent Teams? Hermes Bots, Goal Loops & Kanban Graphs](https://youtu.be/GlW4N7p5KS8) · Wanderloots (Callum), 28 Sep 2026 [^3]: [OpenCode Go](https://opencode.ai/go) · OpenCode, the ten dollars a month plan, checked on 6 Oct 2026