AI copilot skills and tools for the PiDP-1: assembly, debugging, machine operation, display analysis — for Hermes Agent and other agents
Shell
0
68 commits
updated Oct 4, 2026
An AI companion for the PiDP-1 — installs a tutor, guide, and capable copilot on your replica PDP-1. It sits next to you, and can see the display/operate the front panel just like you. Learn how to use the PDP-1 faster, and more enjoyably.
The PiDP-1 is a replica of the original Digital Equipment Corporation
PDP-1 computer.
See the PiDP-1 web site
It is great to have a proper PDP-1 Hacker sit next to you at the front panel. Perhaps having replica Hackers around is just as important as having replicas of their old computers?
I wanted to make an AI companion to the PiDP-1. It is able to vibe-code, but the real purpose is to help people learn how to use the PDP-1 — more quickly, and more enjoyably. Any user, from beginner to advanced.
It is not too hard in 2026. This is a first attempt; I hope it will be picked up and improved. So, feel free to fork, recycle, replace, whatever.
It can install a complete open-source Hermes Agent setup on your PiDP-1 (or any Linux machine). All it needs is an external LLM to connect to. As of November 2026, I'd suggest the cheap, good-enough DeepSeek, but it is up to you. The PDP-1 "problem set" is small enough, and $10 at Deepseek gives you months of playtime. And there is no subscription: when the money runs out, you are off the hook, unless you want more playtime. I like that. The install script will set it up in a few minutes.
This project comes in three independent parts:
Rotating cube |
Tic-tac-toe |
Hopalong radiant attractor(click to zoom!) | Hopalong was ported without human intervention from Javascript in 20 minutes. The agent wrote this page. It depressed me: a sign that the end of democoding is near? |
Admittedly, as a side-effect, the AI becomes a competent PDP-1 coding agent. It was trained doing code reviews of major PDP-1 programs, and it wrote some nice demo code, like the tic-tac-toe game above.
But, personal view: retrocomputing is best left AI-free in that sense. There is no sense of achievement or historical value in letting an AI write an amazing new program on a PDP-1.
Instead, its goal is to be a tutor, click to zoom the screen shot. It can teach you to operate the machine, teach assembly programming, help find bugs. And also, give you an interactive tour through the process of writing code in ET, assembling it, using DDT. Or show you how to work in Lisp. Or take you through a tour of the best graphics demos.
Almost nobody still has an experienced PDP-1 hacker nearby. So it makes sense to have an AI take that role; it will make learning about the PDP-1 much faster and much more fun.
To see some nice demonstrations:
cd /opt
sudo git clone https://github.com/obsolescence/agent-pdp1.git
cd agent-pdp1
./install.sh
[!NOTE] The install script will download and install Hermes for you through their headless install option. The user-friendly way. But Hermes is a bit of a moving target, it depends on - well, lots of Linux dependency hell it seems. So at some point in the future: if the Hermes install through my install script breaks: not a problem. Then, just install Hermes yourself, manually. Easily googled. And rerun the install script afterwards to let it complete the Hermes setup. But this install script is tested to work during 2026 and saved some hassle from a manual install.
From now on, to start working with this: do the usual pdp1control start command, switch the power on (top right of the front panel). Then type hermes --tui any time you want your companion alongside you. For now, use the PiDP-1's GUI setup, the web server alternative is not tested thoroughly yet.
Works on a PiDP-1 (Pi 4/5, 4 GB RAM; untested: 2 GB RAM). If you don't have one, install this after the pidp1 package, on just any Linux laptop — you get a virtual front panel, and everything else stays the same.
The installer presents a menu. Every step explains itself, and every step can be declined:
DRY_RUN=1 ./install.sh simulates the install without changing anything
— handy as a preview, or for testing.
The Hermes step asks for a DeepSeek API key — type/paste it in, or point the installer at a file that holds one (like ~/my-api-key.txt). There is no particular reason to use DeepSeek: it is just, at the time of writing, the cheapest model that is good enough. Hermes runs any model — add others later, things will evolve.
skills/ — core of the project. Readable markdown files with PDP-1 knowledge.
hermes-specific/
install.sh — the installer script. Tested on Raspberry Pi OS and Ubuntu.
None of this needs to be Hermes-specific — and nothing depends on DeepSeek, either. The skills are plain markdown, and most any AI will absorb them; Claude Code, for instance, will happily adopt them. The API that lets an agent work the machine the same way you do is described in the skills files themselves.
As part of this project, the base pidp1 simulator was updated so an AI can share the PiDP-1 with you. Useful for future projects as well. And the PDP-1 simulator gained a massively improved debugger: useful for the AI, so it is not forced to debug over the front panel all the time; useful for human PDP-1 hackers, too. Again, described in the skills files.
We noticed that a self-learning agent like Hermes will, over time, not just learn new things — it will also get caught in its own previous mistakes. Anecdote: when programming Tic-Tac-Toe, it looked online for more example code, found some helpful PDP-8 code, and then got itself confused for a week about the "bug in the subroutine instructions" of the PDP-1…
So, we decided that the self-learning is best kept curated. Instead of
editing the canonical skills, agents are instructed to write what they learn
to their own pdp1-learnings skill file
(~/.hermes/skills/pdp1-learnings/). Read this file in an editor some time —
it is interesting.
We hope people will submit their pdp1-learnings files, so we can use them
to improve the core skills. You can also do that locally:
/opt/agent-pdp1/skills/curate.sh — disable the write protection.pdp1-learnings into the core
skills, and ask it to go through each learning with you, so you can judge
what is useful and what is not. Or, of course, tell Hermes to do it for
itself./opt/agent-pdp1/skills/protect.sh — restore the write protection when you're done.There is also /opt/agent-pdp1/skills/update.sh (or menu option 3 of the
installer): it restores your local core skills to the originals from the
GitHub repo — so there is no risk of breaking anything.
Please use the PiDP-1 Google Group: https://groups.google.com/g/pidp-1
And send your pdp1-learnings files if you want to contribute — through
GitHub, as attachments in the Google Group, or by email. You will have my
email if you are a PiDP-1 user ;-)
/reasoning to medium or low for regular chats, maybe to max or high when you want it to write a program. Tell it to be concise.There is no claim to greatness here. This is experimental, zero copyright or ownership claims — I am just a hobbyist. AI agents will evolve rapidly, so by the time you read this, there may be much better ways to use these skills files. But it is a start: fun, cheap, open source, subscription-free.
We do believe these skills files have some eternal value — they are a thorough knowledge base for anyone, and any AI, to build on. A lot of time was spent getting them effective; improving this curated PDP-1 knowledge is what we think the real value here is.
To close, again, our intention. Preserving PDP-1 legacy is not about vibe coding — this agent can do that, but computer history should be free from AI, IMO.
But...
As people with knowledge of the historical machine start to become scarce, an AI tutor and guide is a very legitimate way of keeping PDP-1 history alive.
Ugh. Lawyers. MIT — see LICENSE. Fork it, recycle it, replace it.
AI copilot skills and tools for the PiDP-1: assembly, debugging, machine operation, display analysis — for Hermes Agent and other agents
Shell
0
68 commits
updated Oct 4, 2026
An AI companion for the PiDP-1 — installs a tutor, guide, and capable copilot on your replica PDP-1. It sits next to you, and can see the display/operate the front panel just like you. Learn how to use the PDP-1 faster, and more enjoyably.
The PiDP-1 is a replica of the original Digital Equipment Corporation
PDP-1 computer.
See the PiDP-1 web site
It is great to have a proper PDP-1 Hacker sit next to you at the front panel. Perhaps having replica Hackers around is just as important as having replicas of their old computers?
I wanted to make an AI companion to the PiDP-1. It is able to vibe-code, but the real purpose is to help people learn how to use the PDP-1 — more quickly, and more enjoyably. Any user, from beginner to advanced.
It is not too hard in 2026. This is a first attempt; I hope it will be picked up and improved. So, feel free to fork, recycle, replace, whatever.
It can install a complete open-source Hermes Agent setup on your PiDP-1 (or any Linux machine). All it needs is an external LLM to connect to. As of November 2026, I'd suggest the cheap, good-enough DeepSeek, but it is up to you. The PDP-1 "problem set" is small enough, and $10 at Deepseek gives you months of playtime. And there is no subscription: when the money runs out, you are off the hook, unless you want more playtime. I like that. The install script will set it up in a few minutes.
This project comes in three independent parts:
Rotating cube |
Tic-tac-toe |
Hopalong radiant attractor(click to zoom!) | Hopalong was ported without human intervention from Javascript in 20 minutes. The agent wrote this page. It depressed me: a sign that the end of democoding is near? |
Admittedly, as a side-effect, the AI becomes a competent PDP-1 coding agent. It was trained doing code reviews of major PDP-1 programs, and it wrote some nice demo code, like the tic-tac-toe game above.
But, personal view: retrocomputing is best left AI-free in that sense. There is no sense of achievement or historical value in letting an AI write an amazing new program on a PDP-1.
Instead, its goal is to be a tutor, click to zoom the screen shot. It can teach you to operate the machine, teach assembly programming, help find bugs. And also, give you an interactive tour through the process of writing code in ET, assembling it, using DDT. Or show you how to work in Lisp. Or take you through a tour of the best graphics demos.
Almost nobody still has an experienced PDP-1 hacker nearby. So it makes sense to have an AI take that role; it will make learning about the PDP-1 much faster and much more fun.
To see some nice demonstrations:
cd /opt
sudo git clone https://github.com/obsolescence/agent-pdp1.git
cd agent-pdp1
./install.sh
[!NOTE] The install script will download and install Hermes for you through their headless install option. The user-friendly way. But Hermes is a bit of a moving target, it depends on - well, lots of Linux dependency hell it seems. So at some point in the future: if the Hermes install through my install script breaks: not a problem. Then, just install Hermes yourself, manually. Easily googled. And rerun the install script afterwards to let it complete the Hermes setup. But this install script is tested to work during 2026 and saved some hassle from a manual install.
From now on, to start working with this: do the usual pdp1control start command, switch the power on (top right of the front panel). Then type hermes --tui any time you want your companion alongside you. For now, use the PiDP-1's GUI setup, the web server alternative is not tested thoroughly yet.
Works on a PiDP-1 (Pi 4/5, 4 GB RAM; untested: 2 GB RAM). If you don't have one, install this after the pidp1 package, on just any Linux laptop — you get a virtual front panel, and everything else stays the same.
The installer presents a menu. Every step explains itself, and every step can be declined:
DRY_RUN=1 ./install.sh simulates the install without changing anything
— handy as a preview, or for testing.
The Hermes step asks for a DeepSeek API key — type/paste it in, or point the installer at a file that holds one (like ~/my-api-key.txt). There is no particular reason to use DeepSeek: it is just, at the time of writing, the cheapest model that is good enough. Hermes runs any model — add others later, things will evolve.
skills/ — core of the project. Readable markdown files with PDP-1 knowledge.
hermes-specific/
install.sh — the installer script. Tested on Raspberry Pi OS and Ubuntu.
None of this needs to be Hermes-specific — and nothing depends on DeepSeek, either. The skills are plain markdown, and most any AI will absorb them; Claude Code, for instance, will happily adopt them. The API that lets an agent work the machine the same way you do is described in the skills files themselves.
As part of this project, the base pidp1 simulator was updated so an AI can share the PiDP-1 with you. Useful for future projects as well. And the PDP-1 simulator gained a massively improved debugger: useful for the AI, so it is not forced to debug over the front panel all the time; useful for human PDP-1 hackers, too. Again, described in the skills files.
We noticed that a self-learning agent like Hermes will, over time, not just learn new things — it will also get caught in its own previous mistakes. Anecdote: when programming Tic-Tac-Toe, it looked online for more example code, found some helpful PDP-8 code, and then got itself confused for a week about the "bug in the subroutine instructions" of the PDP-1…
So, we decided that the self-learning is best kept curated. Instead of
editing the canonical skills, agents are instructed to write what they learn
to their own pdp1-learnings skill file
(~/.hermes/skills/pdp1-learnings/). Read this file in an editor some time —
it is interesting.
We hope people will submit their pdp1-learnings files, so we can use them
to improve the core skills. You can also do that locally:
/opt/agent-pdp1/skills/curate.sh — disable the write protection.pdp1-learnings into the core
skills, and ask it to go through each learning with you, so you can judge
what is useful and what is not. Or, of course, tell Hermes to do it for
itself./opt/agent-pdp1/skills/protect.sh — restore the write protection when you're done.There is also /opt/agent-pdp1/skills/update.sh (or menu option 3 of the
installer): it restores your local core skills to the originals from the
GitHub repo — so there is no risk of breaking anything.
Please use the PiDP-1 Google Group: https://groups.google.com/g/pidp-1
And send your pdp1-learnings files if you want to contribute — through
GitHub, as attachments in the Google Group, or by email. You will have my
email if you are a PiDP-1 user ;-)
/reasoning to medium or low for regular chats, maybe to max or high when you want it to write a program. Tell it to be concise.There is no claim to greatness here. This is experimental, zero copyright or ownership claims — I am just a hobbyist. AI agents will evolve rapidly, so by the time you read this, there may be much better ways to use these skills files. But it is a start: fun, cheap, open source, subscription-free.
We do believe these skills files have some eternal value — they are a thorough knowledge base for anyone, and any AI, to build on. A lot of time was spent getting them effective; improving this curated PDP-1 knowledge is what we think the real value here is.
To close, again, our intention. Preserving PDP-1 legacy is not about vibe coding — this agent can do that, but computer history should be free from AI, IMO.
But...
As people with knowledge of the historical machine start to become scarce, an AI tutor and guide is a very legitimate way of keeping PDP-1 history alive.
Ugh. Lawyers. MIT — see LICENSE. Fork it, recycle it, replace it.