20 days post-Claude Code leak: Did the accidental "open sourcing" actually matter for local devs?

Posted by PaceZealousideal6091@reddit | LocalLLaMA | View on Reddit | 44 comments

Now that its been about 20 days since Claude code source code got leaked, what really came out of it? Sure, we learned some of the inside tricks they use, we understood how much of it is vibecoded, many forks were made... But did it help in any way?

Out of the forks made, I don't even know if any of them work reliably well enough to pay attention to. Did any of the pre-existing popular harnesses actually adopt their parallel tool-calling logic or diffing techniques? I would love to know how if this leaked peeling back the curtain on their orchestration helped anyone here.

I'm asking because, post Qwen 3.6 launch, we're realizing it has become incredibly practical to run highly capable LLMs locally and actually get real work done. With good harnesses and agents, we can execute complex, multi-step workflows we wouldn't have dreamt of even 7-8 months ago, especially on consumer laptops and builds.

Now, we can finally squeeze genuine agentic reasoning into everyday hardware, the model itself is no longer the bottleneck. The harnesses has now the spotlight. I think, now its going to be more about how harnesses are able to make the best out of the model at hand locally.

So, did the Claude Code leak actually give our open-source tools anything to accelerate the evolution? Or it was just a blip that really didn't contribute anything valuable?

[-]

confuseddork24@reddit

I had Claude compare open code, Pi, leaked Claude code, and a custom harness I had made. I've been writing specs for a new harness that I plan to use with local models like Gemma 4 and qwen3.6. It feels like a great time to experiment and iterate on what's already out there to build custom tooling for yourself. I feel like projects like Pi will be more the norm as everyone has different workflows and llms behave very differently based on your workflow and how you interact with it. Having the ability to evolve your harness easily over time seems like the right direction to me.

[-]

my_name_isnt_clever@reddit

Pi coding agent has become by tool of choice for the same reasons, along with hermes-agent as it's the first OpenClaw clone I've used that's actually local friendly out of the box.

[-]

Worried-Squirrel2023@reddit

the biggest takeaway from the leak for me wasn't the code itself, it was seeing how much of the magic is just orchestration. the prompts, the retry logic, the way it chains tool calls. none of it was some breakthrough nobody could replicate. which means the real moat for claude code is the model quality, not the harness. now that qwen 3.6 is closing that gap, the harness becomes something any decent open source project can match. opencode and aider-desk are already most of the way there.

[-]

my_name_isnt_clever@reddit

the real moat for claude code is the model quality, not the harness.

Why did we need a leak to prove this? I've been asking people who swear by CC why they think it's so much better than FOSS alternatives, and haven't had any good answers. Turns out, it's not better. It was just first.

[-]

SourceCodeplz@reddit

The reason you don't know which copy to use is because everyone is build their own.

What we know that is somewhat important is that claude uses many tools, codex less and others more or even less. Also we know that codex does edits with a diff patch while claude does search replace.

So with these said, I think everyone should be making their own harnesses and test what models work better than others. There was a post yesterday where someone used a custom harness and achieved a big boost with a small model locally versus a larger more popular harness.

[-]

PaceZealousideal6091@reddit (OP)

Yeah well, not everyone can build their own. At least not everyone has bandwidth for that. We mortals have just wait for someone trusted to come up with their version and use it when its popular enough and have been vetted enough by others who have the bandwidth to test it out. But, yeah i see that many have made their own. But nothing seems to stick. Nobody is talking it enough.

[-]

Joffie87@reddit

Don't be so down on yourself, I bet you can probably build one too if you want. Ai are pretty good at coding, so your own ability becomes a matter of literacy. I started planning mine last summer with ChatGPT while I was working full time.

[-]

mlhher@reddit

> Ai are pretty good at coding, so your own ability becomes a matter of literacy. I started planning mine last summer with ChatGPT while I was working full time.

AI is pretty good at coding that is true.

Where AI is still relatively miserable is architecture and heavy micro optimization for complex projects, specifically for ones that do not appear often in its training data.

That is why virtually all tools you find whether it is OpenCode, OpenClaw, Claude Code, Codex or some unknown fork of these, all of them are doing more or less the same without anyone attempting any optimization. They just change the feature set, the UI theme and potentially the language and call it a day.

[-]

PaceZealousideal6091@reddit (OP)

That's why I said bandwidth. I can do almost anything I want if I really put my mind to it. It just depends on how time and energy I am willing to put into it. As of now, I don't have time at all for it. And since coding is not my profession, I don't have enough push to do it.

[-]

mlhher@reddit

> There was a post yesterday where someone used a custom harness and achieved a big boost with a small model locally versus a larger more popular harness.

I assume you are talking about me (since I have not seen anyone else barely do these things yet):

I am doing all my dev work in Late (check my GitHub, Profile) at 5GB VRAM with Qwen3.5-35B-A3B at 30t/s often needing no guidance whatsoever anymore and it is significantly faster than all other harnesses because it takes extreme care optimize ruthlessly while everyone else is busy building shiny UIs. Some optimizations aren't even noted in the Readme lol.

But as always do not take my words for granted and try yourself if you are interested (no node_modules, venv bullshit just a single static binary/compile yourself).

[-]

Likeatr3b@reddit

We learned that CC isn’t a work of art, too. I had nothing novel or incredibly well done that experienced devs would have an “ah ha” moment about.

[-]

Far-Low-4705@reddit

regex based mood detector lmao

[-]

philmarcracken@reddit

oh bob saget!

[-]

SourceCodeplz@reddit

Actually it did, some interesting heuristics matches.

[-]

AncientGrief@reddit

off topic:
Codex seems to have problems with creating larger files + with unicode characters (German Umlauts like ä ö ü and even em-dashes seem to have a wrong encoding in the terminal, it always complains about these for me, maybe I am doing something wrong though ... at least on Windows).

Nice to know, that Claude uses diff patches and Codex is not ... maybe that's an issue

[-]

New_Comfortable7240@reddit

Well Roocode and KiloCode have the same issues with unicode for me in my windows laptop, seems more a windows thing, maybe related to using powershell

[-]

AncientGrief@reddit

Yeah I think so too, but Claude doesn't have this problem, wonder what the issue is, maybe a sandbox problem.

[-]

mlhher@reddit

Claude Code is made for cloud models. These people that are telling their models to bloat it, change the UI theme and branding slightly.

A harness has to explicitly think about how to support and build around local inference.

[-]

sine120@reddit

I looked at it, and honestly all it did was confirm CC isn't that impressive. OpenCode seems more elegantly designed with less bloat, but for local stuff, Pi has pretty much been what I've settled on for the smaller system prompt.

[-]

Shot-Buffalo-2603@reddit

Imo this was just a viral “haha anthropic is stupid” meme. The code is typescript which compiles to JS and retains structure meaning at basically anytime you can just open it up and look at a derivative of the real code.

[-]

DelKarasique@reddit

That's bullshit. Compiled js is mostly unreadable after it was uglified, minified and (optionally) obfuscated. Without source maps it is completely useless for reverse engineering.

[-]

Shot-Buffalo-2603@reddit

That’s not bs at all lol my entire job is RE, there are countless publicly available tools the un-minify very well

[-]

DelKarasique@reddit

at basically anytime you can just open it up and look at a derivative of the real code. It’s still cool to see the real code, but anyone who wanted to know what was happening could of just checked at literally anytime

This isn't bs?

"anytime you can just open it up and look at a derivative of the real code" - anytime you can just open it up and look at gibberish which would tell you exactly nothing. Or spend countless hours trying to reverse engineer it back into human readable code.

Yes, reverse engineering is a thing and it's possible but saying "anyone who wanted to know what was happening could just checked it in literally anytime" is just lying.

[-]

NoahFect@reddit

Claude itself is amazing at this.

There's no such thing as a closed-source .NET app now, for instance.

[-]

Jester14@reddit

*could have

[-]

FullstackSensei@reddit

I don't know, I've been doing "real work" locally for over a year. I still don't get all the hype around these tools, because to me it sounds like you're outsourcing the thinking to the LLM. I want to still be in charge of every line that goes into the code base, because I want to be able to maintain it. So, for me, I have been able to do real work even with llama 3 because I always tell the LLM exactly what I want and how I want it. The big difference with the models from the past 6 months or so has been a steady increase in the size and complexity of the task I can give the LLM in one go. Before, in the early days, I had to limit changes to one class/code file at a time. Now, I can ask for a reactor that can span the entire project, though nothing too big or complex, and this is with 200-400B models. Conversely, it used to take 5 minutes to write the prompt for said one class/file refactor, now the prompt can take me 20 minutes or more to write. I haven't noticed any impact from trying different tooling, but again, I keep the model on a tight leesh on what and how to do things.

[-]

draconic_tongue@reddit

this is why I prefer vscode extensions. cli/tuis obscure too much for me

[-]

SkyFeistyLlama8@reddit

Coding harnesses lead to technical debt because no one understands what the hell the LLM spits out, only that it works. Most of the time until it completely blows up one day.

Yeah, SWEs need to realize they're being paid to do the engineering, not to be the monkey between chair and keyboard that approves an LLM's work.

[-]

FullstackSensei@reddit

But, but, but... Garry Tan stopped drinking alcohol to be able to focus 100% on his agents. Just look at his blog, the engine was fully written by LLMs and has 37k lines of code... to load a blog post!!!

[-]

SkyFeistyLlama8@reddit

The number of PRs used to be a success metric. Now it's kilo-lines of vibed code, fed into another LLM to clean up, and then stuffed into another harness to... it's an LLM centipede at this point.

[-]

FullstackSensei@reddit

IIRC, it's still a metric by which contributions are measured by a certain high profile tech billionaire.

Some of the most successful projects I've lead saw almost no code written by me if you looked at the PR history. Nevermind the hundreds of hours helping team members solve technical issues, nor the tens of thousands of lines of text I wrote in requirements, architecture, or fighting management against scope creep...

[-]

kevin_1994@reddit

same. I've been a dev for about 10 years and there's no way in hell i'd ever merge any of this LLM slop code into production. I do find LLMs useful for stuff like debugging SQL syntax query, brainstorming how to optimize a specific set of code, or (probably my #1 most used category of LLM use) writing bash scripts for me

[-]

PaceZealousideal6091@reddit (OP)

Do you also run claude code parallely? do they interfere? is it a rewrite or a fork?

[-]

Limp_Classroom_2645@reddit

No i never used claude code because i never paid them anything

[-]

a_beautiful_rhind@reddit

So, did the Claude Code leak actually give our open-source tools anything to accelerate the evolution?

Lol no.. there's a lot of coding agents that are fully open. Best it could have done is make support of claude itself be better.