Added an extra feature and also made the list more comprehensive.
I added another feature: external SHA sums. Basically, until enough people use this software, most install scripts wont have the SHA256 sum in the format. So, for trusted scripts, the SHA256 sum was generated for each script. This means integrity checking can work on any trusted script, not just the ones which follow the format. I successfully installed Rust on my WSL with this as a demo.

All Done!
Now its fully functional. Install by following the README and use like so:
cursh <url>
or
sudo cursh <url>
if it needs sudo.

Made the script
Now I need to:
1) make it into a pypi package
2) add more trusted sources

Made a database of trusted programs (this is not the only heuristic used to detect unsafe scripts)

OK to prevent confusion: This ship is a terminal AI chat app mainly for terminal help but it can be used for anything. It uses SSH and requires no installation.
DO NOT USE THE demo as its for the older ship ( regular chat app ).
To use it, type ssh tanjim.hackclub.app -p 5022 in your terminal (any username/password)
This is similar to GH Copilot CLI (gh copilot suggest <what command you want>) but requires no installation, uses better models and has context.
The previous ship was a regular chat app that used Github Models to provide a wider selection of models for free. I hope that cleared everything up.

Alright finally made this work! Type ssh tanjim.hackclub.app -p 5022 (any username/password) to try it out. This will be useful if i forget a command/just want to know how to do a thing in the terminal.

Ok, because of SOM server issues this devlog has to cover a lot. I am making a terminal version of LLM dump that uses SSH. I made the SSH server script, now I have to put it in the nest server. Isnt this so cool!

Ok, not all features I initially said are available :(
For now, every custom command component except use_cmd does not work. This means that the item can be as complicated as an item given with a custom /give command (including data components/NBT), but with the exception that commands can run on use.

And also, it has to be an item, not a block or entity.

Examples:
- Amethyst shard that gives diamonds on use: yes
- Blaze rod that gives speed on use: yes
- sword that deals 20 damage: yes
- armor that increases your health by 50%: yes, but you may have to explicitly say to do this using only use_cmd and attribute modifiers
- sword with lightning: sorry no
- sword that deals 7 damage and gives speed on rightclick : yes
- sword that deals poison to enemy : no

YOOO FINALLY VERIFIED IN MODRINTH!
Go download the mod from https://modrinth.com/mod/idea_anvil and use the /idea command to get items! hooray!

YOOO IT WORKS LETS GOOOOOOOOO
It finally works. Use /idea <prompt> or /ai <prompt> to get items. After submitting to Hackcraft, I will add a feature where you can export your items to a new datapack/mod. I know this may ruin modrinth with AI-generated mods, but inevitably, someone else will do it if I don't.

The ai works, it gives me the correct item, but some weird error is happening on the use_cmd thingy
This diamond giving shard was given by AI.
I typed the command /idea a shard which gives diamonds
The AI did a perfect job (according to /data) but now something is happening in my mod? gotta debug again :(

(note: i accidentally deleted it to submit to another YSWS before realizing its double dippable, please please accept, like this fixes a REAL problem many people had + the package somehow has 2.59k downloads)

Added comparision + badges to README, made final touches

I found out the project somehow has 2.59k downloads? I added badges to flex that as well. If you are wondering why not use MSYS2 (or git bash)/Coreutils/Busybox/WSL, you may want to check the comparision out.

Added all needed utilities, fixed powershell problems

I added a section to the installer that removes conflicting aliases and saves the config. When uninstalling, it restores the previous config.

Made readme better

Now its less confusing

Added more utilities

Utilities like sha256sum and ps have been added. ls cat etc still have problems

Added some utilities, PowerShell problems

Only a few utils have been added for now. And the default PowerShell aliases cause problems for commands like cat ls cp mv etc so you have to type .exe at the end for now.

Made cool logo

Typed 'mkcd' in Roboto Mono with the text cursor visible then put a terminal icon. Isn't this cool?

YOOOOOO FINALLY THE CUSTOM ITEM SYSTEM WORKS! LETS GOOOOO!
This diamond giving shard (gives diamonds on right click) can be gotten with the command:
/give @s the_idea_anvil:custom_item[minecraft:item_model= "the_idea_anvil:amethyst_shard_lime", the_idea_anvil:use_cmd="give @s diamond", minecraft:item_name="Diamond giving shard"]
Note that while the screenshot loooks like the player needs permission, the command is actually exec'd with op level 2 (enough for give but not enough for op)
This can make almost any datapack item. The only exception is when you need to add new commands to run every tick. This is almost mitigated because you can still specify the_idea_anvil:inventory_tick_cmd which runs every tick in the inventory, but when its outside the inventory (dropped) it wont run. So, there are almost no exceptions. Also, you can run multiple commands with #/ as seperator (but not ##/), example: give @s diamond #/say I got a diamond!
Cmd component list:
the_idea_anvil:...
- use_cmd : #s gets replaced with self UUID
- post_hit_cmd : when hitting an entity, #s is attacker UUID #t is target UUID, for example execute at #t run summon lightning_bolt
- post_mine_cmd: #s is miner UUID, #x,#y,#z are coords of the mined block, #p is #x #y #z space-concatenated
-inventory_tick_cmd: #s is self UUID
- use_on_block_cmd: #s is self UUID, #x #y #z are coords, #p = #x #y #z space seperated

Added the temporary texture system for now.

re-added the texture system, for example here is what would happen if you asked for a grass colored (amethyst) shard. some item textures are recolored to 32 different colors (including the 16 wool colors, materials like diamond, etc) then added to a fixed texture list. the AI can use anything from this colored list or default textures. this has 1732 textures for the colored ones plus however many default textures. i will improve this to let the AI generate something but the problem is free image generator APIs have low ratelimits (ai.hackclub.com has infinite requests for LLMs free but not image)

i accidentally deleted my entire project.
thanks to github a lot of the changes (not most) were backed up. but i almost made everything and a lot of things got deleted. well i am not giving up since this 8h was mostly spent on experimenting with the MC apis and stuff until i found a working solution so it will take way less than this time (1-3h). well i have more hours of things to log to SOM

Done thinking about the idea and made a logo.
Also made the structure of the mod.
How it will work: 1) You use a block called the Idea Anvil and ask it to make something
2) It gives you an idea item.
3) You can bring it to the crafting table and use it to craft your item. Your item's recipe will be decided by the AI as well (If you are in creative, you can get custom items from the creative menu if you have the idea in your inventory)

Fixed some bugs. Starting to add external APIs.

Added model tags so you can actually know what model to use. Added support for even more models. Made ratelimits clearer.

I finally added web search. It works now. It uses DDGS internally in the Azure backend. I tried using the official duckduckgo api, but it does not exist anymore.

Added more websearch stuff, websearch still not complete. I made the button and the function, now adding toolcalls.

Devlog for Hack Club

Aug 18, 2025

I was making this project before hack club. So I am just writing the current fetures I added to the project.

Idea

Github has this free and unknown API called Github Models. It gives access to every model ever except Anthropic. Now, unlike other free APIs, you dont need to use APi keys and any app you sign in with github can use it.
So if I make a website that uses this API, it will seem easy to use (you dont need to go register for an API key) and it will seem like Github is for auth while its the entire thing this project depends on.
You can use literally every model. Every single one. But not Anthropic. I am not sure why Github Models does not support Anthropic, but it does not and thats how it is.
You can use GPT-5 Chat, GPT-5 Mini, GPT-5 Nano, GPT-5, 4.1, etc with awesome ratelimits. Check out every model at https://github.com/marketplace?type=models.
Features currently:
- Sign in with Github
- Use any model you want
- Multi turn conversations
- Editing and branching
- Saving conversations
- Title generation
- Custom system instructions
Planned:
- Web search tool
- Python Code execution tool (using pyodide)
- Custom tools +marketplace (example: you can make a tool that allows it to use wolfram alpha, then upload it to the marketplace. Or just download it because someone already made it)
- Multimodal file uploads
- Image/Video gen
- Custom APIs
- Artifacts (if the AI makes an HTML/CSS/JS app you can use it directly in LLM Dump)
- Coding Agent with unix terminal and file system, with support for a virtual network (so if it hosts something on localhost you can access it, and it can access the internet and its own localhost as well).