Custom v95 post-bb MapleStory client.

[LastBattle]

What would you do when there is finally have a 1:1 custom client that works just like the official client, with all of its quirks, and bugs? Fully AI-replicated with full access to the source code.
What sort of deep customisation to the game to your preference?
I have this idea for an awfully long time, but often always time constrained in my personal life.

Well, over the past 4 months I've been working on exactly just that, since the advent of claude-opus-4.6 in December 2025, something that previously can only dream of has finally been fully unlocked.
the ability to summon infinite amount of shadow devs at will to tackle at any problems you cant solve alone.

The constraints in a post-AI dev world are no longer the number of devs that knows the internal workings of the MapleStory client (reverse engineering + wz) or human-time. At the moment the only major bottleneck I face is really only the lack of tokens (energy/ money constrained).
In order to automate these at scale, it is all about the quality of your .md plans, and MCP server (IDA-pro, wz/ img) that the agent is able to work autonomously 24/7 while you sleep.

I expect that within a year as AI gets 10x better, cost would equally be falling by about 90%, with the same workflow, and it will be possible to clone any version of MapleStory client with as little as $100 in subscription (a single codex or claude-code).
Picture this: any client, any version. as long as you have the MapleStory.exe and .wz archived, you will be able to replicate it. You do not need any source code, whatsoever.

Even at the current phase of AI, in the development of this, I've personally only written less than 100 lines of code by myself over the past 4 months. The job as a software engineer is largely over, everyone is either a solo founder / project manager focusing solely on the architecture/ goals/ plans or gets eliminated from the field.

2026-04-03

In the meantime, if you'd like to follow the development, check out the staging branch of Harepacker-resurrected:

To view the content, you need to sign in or register

It is all open source.
C# .NET 10.0 base (wouldnt matter much a year from now anyway as porting to Rust / wasm would just be as simple as a few prompts)

Tooling/ plans if you are interested in trying it out yourself.

• WzImg-MCP-Server
  To view the content, you need to sign in or register
• ida-pro-mcp (requires IDA-Pro 8.0 and above, DM if you need)
  
  To view the content, you need to sign in or register
• Plans:
  To view the content, you need to sign in or register
• Agent Skills (
  To view the content, you need to sign in or register
  ) to be used with gpt-5.4 with medium thinking effort.
  
  To view the content, you need to sign in or register

It will be a stand-alone client eventually, detached from HaCreator, along with the full ability to connect to a v95 server.
Keep an eye on this thread

 

[PacketRoute]

Looks really cool! nice to see you building on your WzLib/HaSuite stack
LLMs really accelerates RE speed, feels like we're getting closer to clean client reconstructions, at least for v95/v176 (using the KMST 1029 pdb)

 

[LastBattle]

Looks really cool! nice to see you building on your WzLib/HaSuite stack
LLMs really accelerates RE speed, feels like we're getting closer to clean client reconstructions, at least for v95/v176 (using the KMST 1029 pdb)

Its a really fun side-project I must say, to keep up with the AI space.

v95/KMST 1029 or those with pdb initially, then you could move on to all adjacent versions such as v101 -> v105 -> v110. Gradually fill that entire version gap by automatically identifying the functions, naming, variables, etc.

When there are no longer any constraints on the cost of tokens, all of these will be possible.

-------------------------------------------------------------

included the

• Agent Skills (
  To view the content, you need to sign in or register
  ) document, to be used with gpt-5.4 with medium thinking effort.
  
  To view the content, you need to sign in or register

 

[sh00ter999]

Very interesting project.
Your video showcase seem broken, is that intentional?
How much $$$ did you have to spend over the duration?
Did you let it work non-stop over these four months or was it a on-and-off type of process?

The resources you linked are quite neat. Honestly, I was sure that any sort of reversing was still out of reach but the ida-mcp looks cool.
It seems to burn through tokens like a hot knife though?

Exciting times ahead for sure.

 

[LastBattle]

Very interesting project.
Your video showcase seem broken, is that intentional?
How much $$$ did you have to spend over the duration?
Did you let it work non-stop over these four months or was it a on-and-off type of process?

The resources you linked are quite neat. Honestly, I was sure that any sort of reversing was still out of reach but the ida-mcp looks cool.
It seems to burn through tokens like a hot knife though?

Exciting times ahead for sure.

I've reuploaded the video, it seems that the links only last for 2 days for unregistered account.


Its estimated that I have used around 10 billion tokens thus far, running on early generous limits of Anthropic's Claude-Code (opus-4.5) seemly all-you-can-eat buffet in early january. They have since cracked down on the superpower users. I'm now relying on OpenAI's codex (gpt-5.3-codex, gpt-5.4) ChatGPT Pro subscription.
In terms of subscription cost personally, perhaps around $800-1000 in total running 2 subscriptions a month and exhausting all of what I am able to use. As far as I am aware, it currently cost somewhere north of 10-25x to provide those services (datacenter + GPU capex), so those companies combined might have lost way over $50,000. Its all VC-subsidised anyway to get you and corporations hooked into AI-workflow before cutting off the free opiums. Until the AI party ends (like Uber and DoorDash in its early days), abuse it while you can.

Its working non-stop autonomously for as long as I still have any sort of 5 hour / weekly quota remaining, much of the pipeline on my side has already been fully automated. The one and only limiting factor is tokens, give me infinite tokens and I can do infinite MapleStory clients.

If we try to project a year out from with the double-exponential AI progress when the current level of intelligence (which is just about enough to do reverse engineering work) falls by 10x-100x, we should be able to create a client for as low as $100-200 each. Its within reach it all depends on giving it enough context, and build good harness (MCP server) around it. And perhaps some level of coordination problem to solve (i.e how do you have 10-20 agents working concurrently while not conflicting with each other's work; the same way humans do!)

You may reference the leaked Claude-Code source code

To view the content, you need to sign in or register

, or the open source codex

To view the content, you need to sign in or register

on ways of doing this effectively, including the MCP + skills.md I have provided in this thread.
maybe also read up on openai developer blog

To view the content, you need to sign in or register

and most importantly, get on x

To view the content, you need to sign in or register

its the only social network with a great concentration of where all AI talents are at, everywhere else you will be behind by at least 6 months in this frontier.
The average person using AI these days on a free chatgpt account still thinks its a dumb chatbot from 2023 when much of the white collar class will already be eliminated a year from now.

All of the current open source AI models such as qwen, Z.AI GLM, llama, gemma does not exhibit sufficient enough intelligence to do this in any meaningful way, its simply making a mess.

 

[sh00ter999]

Awesome. Appreciate the detailed response

All of the current open source AI models such as qwen, Z.AI GLM, llama, gemma does not exhibit sufficient enough intelligence to do this in any meaningful way, its simply making a mess.

Oh yeah, I've made plenty of experience with that; hence why I didn't initially have much faith when testing 4.6 and Codex. I'm quite happy being able to send them off to do minor tweaks in larger codebases according to personal preferences (that actually end up working most of the time). I've also heard the same projection that within a year (or two) from now, we might be looking at a drastic cost reduction for models with increased capabilities.

Hopefully we get there without any large scale disruption. The Claude Mythos news sounded a bit concerning with it finding ancient exploits out in the wild. If that went rapant, who knows what could happen

Grok Code/xAI is also evolving rapidly. It's fast and cheap(er) but realistically still not up to par with the other top contenders I think. The most fascinating thing here is observing how their orbital ambitions will fare and if "space datacenters" will ever be a thing. Certainly not a near-term type of deal, but who knows! One of my main concerns is that pricing won't budge and that they will continue to identify price points that are juuuust within a customer's pain tolerance zone, even though they could easily offer better deals. Look at phone carriers for instance. We may get larger data plans and speeds compared to one or two decades ago but the price is mainly the same - which on paper is still a net gain for the end user. But then there are these arbitrary limitations, such as if you do not use up all your bandwidth within a random 30 day time period, you lose it all and you gotta keep the money flowing or else -- stating the obvious but this has always been irritating. Token subscriptions may likely follow the same path but one can hope and dream, eh?

Sorry, don't mean to derail any further. Had to get these ponderings off my chest. Your client rewrite project is fascinating. Definitely one of my dream projects as well since I'd say it's one of the largest hurdles holding back a game with a great foundation.
One had to wait many countless years for official 1080p support, 64-bit, minor performance tweaks but all of that is barely scratching the surface at this point. I'm sure it's very difficult but certainly not impossible, especially if you had money, time, resources, and the full source code at your disposal as N*xon does.

Thanks again for the showcase. I'm definitely curious to see where you ultimately land.

 

[LastBattle]

Awesome. Appreciate the detailed response


Oh yeah, I've made plenty of experience with that; hence why I didn't initially have much faith when testing 4.6 and Codex. I'm quite happy being able to send them off to do minor tweaks in larger codebases according to personal preferences (that actually end up working most of the time). I've also heard the same projection that within a year (or two) from now, we might be looking at a drastic cost reduction for models with increased capabilities.

Hopefully we get there without any large scale disruption. The Claude Mythos news sounded a bit concerning with it finding ancient exploits out in the wild. If that went rapant, who knows what could happen

Grok Code/xAI is also evolving rapidly. It's fast and cheap(er) but realistically still not up to par with the other top contenders I think. The most fascinating thing here is observing how their orbital ambitions will fare and if "space datacenters" will ever be a thing. Certainly not a near-term type of deal, but who knows! One of my main concerns is that pricing won't budge and that they will continue to identify price points that are juuuust within a customer's pain tolerance zone, even though they could easily offer better deals. Look at phone carriers for instance. We may get larger data plans and speeds compared to one or two decades ago but the price is mainly the same - which on paper is still a net gain for the end user. But then there are these arbitrary limitations, such as if you do not use up all your bandwidth within a random 30 day time period, you lose it all and you gotta keep the money flowing or else -- stating the obvious but this has always been irritating. Token subscriptions may likely follow the same path but one can hope and dream, eh?

Sorry, don't mean to derail any further. Had to get these ponderings off my chest. Your client rewrite project is fascinating. Definitely one of my dream projects as well since I'd say it's one of the largest hurdles holding back a game with a great foundation.
One had to wait many countless years for official 1080p support, 64-bit, minor performance tweaks but all of that is barely scratching the surface at this point. I'm sure it's very difficult but certainly not impossible, especially if you had money, time, resources, and the full source code at your disposal as N*xon does.

Thanks again for the showcase. I'm definitely curious to see where you ultimately land.

The same was the case with gpt-4, screaming *agi agi, terminator* and after a week after everyone got used to it and just stopped bothering shl*t. gpt-5.3-codex was gate-kept for a couple of weeks before it got broadly released due to security concerns. In my opinion, Anthropic do have an incentive to hype up safety to get a regulatory moat and monopoly within congress, I would just ignore much of what it says.
2 years ago I did have concerns during the Biden administration that AI could be kept strictly away from the public use just as the Mannhattan Project during WWII, and hoarded a couple of GPUs myself, Today, it is no longer the case. Trump administration changed all of that, and now the cat is already fully out of the bag, along with competitive pressure from China, theres nowhere but only to accelerate.

Grok Code/xAI is certainly interesting, as models continue getting commoditized the next bottleneck becomes solely on energy, materials constraint and the ability to buildout in the physical world. I wouldn't fade elon, having proved that again and again with SpaceX, Tesla.
Pricing in my opinion probably depends on when the AI bubble *does break*, forcing all companies to tighten their finances.
A good measure is if you are using more tokens when models get better, if you do not need more it means much of the buildout gets increasingly stretched thin and by definition that becomes a bubble. That has not been the case so far, more buildout just results in even more usage today. (Jevons Paradox)
Also from geopolitics game theory perspective, the USA is currently locked into the AI war with China, its hard to see any rational administration allowing a recession at this point of time regardless of how overheated AI may be. Yeah, really an insane era we live in

Its hard for corporations to move fast relative to an individual, all of that layers of management, departments, security, etc.
It was a fun project on my side anyway, keep a lookout! Thanks.

 

[coolgepds]

May I ask, is this the client source code?
Which version of the client is used?
And which version is the server?
I'd like to give it a test run!

 

[LastBattle]

May I ask, is this the client source code?
Which version of the client is used?
And which version is the server?
I'd like to give it a test run!

An AI-recreation of the v95 client, not the client’s source code.
It’s still a work in progress, check out the link in the thread to follow the development.

 

[klaiu228]

Hii! It's been a while. I have sent you a DM in case you did not saw it. I have a question and I wanted to ask it here, if that's okay!

So, I wanted to know how well does the model is behaving recently? Did you had any major problems running your Maple Project (like some issues with token resets or quality drops in the model's performance) if you are still using it? I saw some people having some trouble using Codex and I wanted to hear from your progress, since you are putting lots of money and effort on it! I hope you are having a great time developing this project, I'm having a great time following the development of it!

 

[heyweidao]

I'm also doing the same thing !

 

[coolgepds]

I'm also doing the same thing !

View attachment 302175

你的QQ号是？一起研究

 

[heyweidao]

你的QQ号是？一起研究

my QQ: 3328256797

 

[LastBattle]

I'm also doing the same thing !

thats cool! keep it up
--------------------------------------------

btw I'm putting a pause for a bit as I've largely saturated gpt-5.5's ability to do this with just client assembly and wz view of the game, perhaps until there's a new model or better harness (i.e vision, ability to play and test the game).
It been going in circles, burning tokens without much further progress.

I will come back to this in a few months or so, time is on our side to just wait it out without having to do this manually. its all a tech issue. gpt-5.6 sol, mythos etc

 

[LastBattle]

Heres some teaser in the meantime.

Spoiler

and some more

Spoiler

getting the last 5% to perfection is always the hardest problem to solve.

 

[mchronos]

im making somethign similar pure ts browser not wasm

 

[LukasMuler]

will you post the client ?

 

[shinhwe]

will you post the client ?

it's open source, and the source code github has been provided.

 

[LastBattle]

I've just merged it into the main branch of HaCreator, you may compile it yourself to run it for now. Its stable and fast enough for a beta-release.

To view the content, you need to sign in or register

Recommend using post-big bang MapleStory for the best compatibility. (v95 to v260).

 

[LastBattle]

Merged, and available here in the latest version of HaCreator map simulator:

To view the content, you need to sign in or register