The factory is only as valuable as its products. The real question is always: how good are these products?
I see a SQLite clone on that list. How does it compare? What are the trade-offs? What did the factory do, vs a plain Claude prompt, vs just using SQLite?
Relatedly, how novel are they? Maintaining a "clone of X" is likely easier than building something with new core ideas, especially if X or X-like things are already open source.
TFA does assign every entry a "complexity" and "innovation" score and explains the rationale behind each if you click on it. The scores can be debated but at least there is some consideration given to the aspects raised in this thread.
One thing that provides some superficial confidence in the author's self-assessed scores is that most of the innovation scores are 1/5.
I'll add: why should I use the clone instead of the original? Cloning is as old as humanity, but usually the clones have some competitive reason to exist: price, availability or a minor value-add like a different color.
All of these have a competitive reason to exist. That was an explicit filter to something making it on the list.
At the very least, it ditches an external dependency, provides better developer experience, is more performant, adds features, or fills an unserved niche. Usually several of those.
Edit: forgot to add that a lot of .NET projects are moving to commercial licenses. See ImageSharp, for example.
Absolutely not clear in a site that looks like a vibe-dump. I think it would be a good idea to hand write the value proposition for these libraries very clearly.
I don't know why this was flagged... whoever wants to play the card of "llm so is bad" they need to do at least some kind of research into it, if not it's just randomly dismissing things just because it has this one word.
I have not done that due dilligence myself, but I know increased output is possible, and the better you know the domain and how to set the "harness" around it the easier is to control and produce the output.
When the postman thing happened I pulled the trigger from doing the replacement for https://github.com/alainnothere/mailBoy and 4 or 5 hours later I was using it loading the same collection to do what I need...
I used it, validate error and have done updates to it.. so I can vouch for it...
If he in any way can validate and make sure the quality is up to production "to be sold" and clients are using it it goes to show you can do it... yes one person shipping this libraries...
Love this. I think we should all try to be more ambitious and build more of the tings we use. It's fun, and you have so much more control over what you can build.
I did this myself and built my own Vercel + Supabase stack on Hetzner. Managed to save a lot on bills, but more importantly, having that low-level control let me add things like auth, db, kv, domain purchasing directly from Claude Code, search indexing for websites, analytics, etc. And I can run as many projects as I can fit into a 4GB RAM box.
Does this mean it's going to replace Vercel? Probably not. But it's fun!
I think LLMs exacerbated an existing ongoing trend - people keep making new libraries and middleware, instead of end products. i.e. Use this tool to make your thing that you actually wanted to do, vs. something that does it directly (e.g. yt-dlp).
I'd actually love to see end products shown here - e.g. a game, a display of real data, a terminal cli that does something useful directly. But I'm not going to pick among the 1000 competing vibe coded library for my own projects. I'll stick with the ones started pre-2024 or LLM myself.
It feels like while that's 'true', there's also an explosion of end products as well. Depending on the communities you hang out in, you might start getting sick of the nth gridfinity box creator or solar battery layout system...
On the other hand, we're also getting actually useful stuff as well. Solidcraft is one vibecoded app that I'm genuinely impressed by. It's gone from interesting toy, to actually useful in a few short weeks.
The cost of creation has dropped to 1/10x what it used to be. The cost of curation is now what's increasing. But that's always been true of any human output. Vanity presses, high-school blair-witch knockoffs, fan-fiction, myspace buffy fan pages. Software is now just another output that is no longer out of reach of the regular joe. And there will be a lot of crap made. But at the same time, you'll still get gems.
As long as I can directly use it for useful purpose or an actual problem that I have, I don't mind. But those are lot harder to come by and usually already occupied by existing free and open-sourced product.
There is a gap in the C# developer ecosystem for native AOT compatible libraries doing these kinds of high-level operations (GPU/Web/Audio/PDF/LLM), that don't rely on shim DLLs (aka DLLs written in C/C++), so I am intrigued to see where this goes.
For as long as software engineering has existed, there have been prolific developers and there have been slow developers and every developer in between. Success has never been a product of volume. Some guy producing 30 "complex open source libraries" by hand in 2020 or with a "software factory" in 2026 doesn't mean anything. @sindresorhus published >1k packages long before code could be generated. Software is a means to an end, generating 30 "complex open source libraries" for no end changes nothing. If anything, these software factories are showing just how worthless most code is. Certainly, a "software factory" can be an accelerant towards building software that achieves some end, but it cannot magic an end out of nothing, it cannot make software valuable or useful, purpose must come from people.
> @sindresorhus published >1k packages long before code could be generated.
Interesting that you mention him, I was just thinking of him when I started reading your comment. Back in the day, he was famous for publishing lots of single-function npm packages like is-ip, is-regexp, copy-file, memoize, etc. Interestingly, he has now pivoted to mass-producing single-function macOS and iOS apps using AI. I guess some things never change.
Distribution has always been the bottleneck but that's becoming even more true now.
I have a one-person product and I'm exploring how to out-distribute incumbents who already have reach. I don't know if speed to market is a differentiator but cost surely is. Curious how others might be thinking about this.
>> I have a one-person product and I'm exploring how to out-distribute incumbents who already have reach. I don't know if speed to market is a differentiator but cost surely is. Curious how others might be thinking about this.
I have a one-person product that AFICT has no incumbents. It's an old idea of mine that AI could turn into an app in a weekend. Problem is if I finally make this, how do I get it out there before the clones come? I assume there are multiple players scanning the app stores for new things to clone and put under their already popular brand. One could get scooped even with a new idea. Am I off base with that concern? How to launch?
It's only a problem if the clones have solved distribution - which I don't believe they would have. Else they would have created a product the hard way without AI already.
I think if you can figure out how to find your customers and put your product in their hands, then AI is a true enabler. If you can't, AI doesn't help at all.
It’s not that frightening. Paying for tokens to keep this up and make sure it’s patched is boring work that folks would gladly outsource.
Also not sure why I’d trust one dude to be able to effectively manage this much context, or find all the bugs and issues. Or even patch them. I certainly wouldn’t trust myself to do it - and I’m an experienced engineer also doing a lot of work with AI.
Also don’t forget the luxury that is blaming other people for outages. It’s worth a lot of money.
The invisible hand is a metaphor inspired by the Scottish economist and moral philosopher Adam Smith that describes the incentives which free markets sometimes create for self-interested people to accidentally act in the public interest, even when this is not something they intend.
> Recursive AI is Kurtis Welch and the fleet of AI agents he directs. You work with one senior engineer who owns the result, and you get the output of a full team.
None are dumb clones. The ones that are the closest to "clones" are faster than competing libraries and have more developer friendly APIs and don't have external dependencies.
Interestingly, these libraries aren't the full factory output. It's just the part that's easiest to share. I can't exactly share production client code.
In 30 days of shipmas, this org has recreated libraries that already exist.
Also, if you're going to just recreate open-source projects, shipping 1 a day is not impressive at all.
Anyone can rip off 7zip, imagemagick, lucene, spacy, whisper etc and call it innovation
>Full list to save people a click
>OnePdf — Read, repair, edit, sign and fill PDFs, and write PDF/A and accessible PDF/UA.
>OneDocument — Read, edit, recalculate and write Word, Excel and PowerPoint files, and render them to PDF.
>OneHtml — Parse and edit HTML like a browser, and check accessibility without one.
>OneRender — Lay out HTML and CSS and render them to pixels or tagged PDF without a browser.
>OneFont — Load fonts and shape text in any script, with bidirectional layout and line breaking.
>OneImage — Decode, edit, convert and resize images, from PNG and JPEG to layered PSD.
>OneLite — Embed a SQLite-compatible database that reads and writes the same files as sqlite3.
>OneIndex (preview) — Run full-text, vector and hybrid search in-process or in the browser.
>OneArchive — Compress and unpack ZIP, 7z, tar and other archives, with identical output across platforms.
>OneMail — Read, write and convert mail stores, including Outlook PST, and send and fetch mail.
>OneCrypto (preview) — Load keys and certificates, validate certificate paths, and sign and encrypt with OpenPGP.
>oneyaml — Load, dump and edit YAML in place in Rust, Go and C#, preserving comments.
>OneContact — Find, parse and normalize phone numbers, emails, URLs, addresses and names.
>OneDiff — Diff and merge text like Git, and redline changes between digital or scanned PDFs.
>OneModel — Run and train ONNX models in pure C# on CPU, GPU or in the browser.
>OneNlp — Tokenize, tag, parse and extract entities from text, and train models in C#.
>OneSpeech — Transcribe speech, add timestamps and identify speakers with Whisper and other open models in .NET.
>OneOcr — Read text, tables and invoice fields from scans and PDFs.
>OneLlm — Call Anthropic, OpenAI and Gemini models through one API and run agents defined in files.
>OneModelContext — Serve tools to AI agents over MCP or connect to MCP servers with OAuth.
>OneEcma — Run JavaScript inside .NET with execution limits and clear error reports.
>OneShader — Compile HLSL to Vulkan, Metal and WebGPU shaders as part of dotnet build.
>OneGpu — Run compute and ML kernels on Vulkan, Metal or WebGPU from pure C#.
>OneAudio — Decode, encode, mix and resample audio, measure loudness and play audio on devices.
>OneMedia — Remux, trim, tag, repair and package video and audio without re-encoding.
>OneGame — Let AI agents build, run, test, inspect and change games through commands without an editor.
>OneMath — Compute sin, exp, pow and more with correct rounding and bit-identical results across machines.
>OnePath — Find paths on grids, graphs and navmeshes, build navmeshes and steer crowds.
>OnePhysics — Simulate 2D and 3D rigid bodies, cloth, soft bodies, fluids and vehicles with bit-identical results across machines.
>OneScope — Scrub, fork and compare live or recorded runs of deterministic simulations.
Are any of these value add? I can create OneDiff in one line of code in .bashrc `alias OneDiff=diff`
Hilariously, none of these say what they do differently. OneModelContext, how is that different than MCP? It's description is literally a description of MCP, but it says it connects to MCPs. So what does it actually do? Reading the repo, it says it is an SDK for MCP, but then it also says it's an API for MCP. Which is it?
What value are these things if you can't even get the readme to agree with what it is? Why should I trust any of it to work correctly?
Additionally, the oneyaml's github link goes to this company's github profile, not a repo and clicking the link for oneyaml in the profile sends you back to this website. If you can't even link to the repo correctly, what are the odds of anything working correctly?
The factory is only as valuable as its products. The real question is always: how good are these products?
I see a SQLite clone on that list. How does it compare? What are the trade-offs? What did the factory do, vs a plain Claude prompt, vs just using SQLite?
Relatedly, how novel are they? Maintaining a "clone of X" is likely easier than building something with new core ideas, especially if X or X-like things are already open source.
TFA does assign every entry a "complexity" and "innovation" score and explains the rationale behind each if you click on it. The scores can be debated but at least there is some consideration given to the aspects raised in this thread.
One thing that provides some superficial confidence in the author's self-assessed scores is that most of the innovation scores are 1/5.
I'll add: why should I use the clone instead of the original? Cloning is as old as humanity, but usually the clones have some competitive reason to exist: price, availability or a minor value-add like a different color.
All of these have a competitive reason to exist. That was an explicit filter to something making it on the list.
At the very least, it ditches an external dependency, provides better developer experience, is more performant, adds features, or fills an unserved niche. Usually several of those.
Edit: forgot to add that a lot of .NET projects are moving to commercial licenses. See ImageSharp, for example.
Absolutely not clear in a site that looks like a vibe-dump. I think it would be a good idea to hand write the value proposition for these libraries very clearly.
I don't know why this was flagged... whoever wants to play the card of "llm so is bad" they need to do at least some kind of research into it, if not it's just randomly dismissing things just because it has this one word.
I have not done that due dilligence myself, but I know increased output is possible, and the better you know the domain and how to set the "harness" around it the easier is to control and produce the output.
When the postman thing happened I pulled the trigger from doing the replacement for https://github.com/alainnothere/mailBoy and 4 or 5 hours later I was using it loading the same collection to do what I need...
I used it, validate error and have done updates to it.. so I can vouch for it...
If he in any way can validate and make sure the quality is up to production "to be sold" and clients are using it it goes to show you can do it... yes one person shipping this libraries...
Love this. I think we should all try to be more ambitious and build more of the tings we use. It's fun, and you have so much more control over what you can build.
I did this myself and built my own Vercel + Supabase stack on Hetzner. Managed to save a lot on bills, but more importantly, having that low-level control let me add things like auth, db, kv, domain purchasing directly from Claude Code, search indexing for websites, analytics, etc. And I can run as many projects as I can fit into a 4GB RAM box.
Does this mean it's going to replace Vercel? Probably not. But it's fun!
You can check it out, its open source
https://shiptiffin.com/
I think LLMs exacerbated an existing ongoing trend - people keep making new libraries and middleware, instead of end products. i.e. Use this tool to make your thing that you actually wanted to do, vs. something that does it directly (e.g. yt-dlp).
I'd actually love to see end products shown here - e.g. a game, a display of real data, a terminal cli that does something useful directly. But I'm not going to pick among the 1000 competing vibe coded library for my own projects. I'll stick with the ones started pre-2024 or LLM myself.
It feels like while that's 'true', there's also an explosion of end products as well. Depending on the communities you hang out in, you might start getting sick of the nth gridfinity box creator or solar battery layout system...
On the other hand, we're also getting actually useful stuff as well. Solidcraft is one vibecoded app that I'm genuinely impressed by. It's gone from interesting toy, to actually useful in a few short weeks.
The cost of creation has dropped to 1/10x what it used to be. The cost of curation is now what's increasing. But that's always been true of any human output. Vanity presses, high-school blair-witch knockoffs, fan-fiction, myspace buffy fan pages. Software is now just another output that is no longer out of reach of the regular joe. And there will be a lot of crap made. But at the same time, you'll still get gems.
As long as I can directly use it for useful purpose or an actual problem that I have, I don't mind. But those are lot harder to come by and usually already occupied by existing free and open-sourced product.
Yeah everyone's busy reinventing the wheel without checking if people want to commute on a unicycle.
There is a gap in the C# developer ecosystem for native AOT compatible libraries doing these kinds of high-level operations (GPU/Web/Audio/PDF/LLM), that don't rely on shim DLLs (aka DLLs written in C/C++), so I am intrigued to see where this goes.
Hard for it to write simple libraries, LLMs tend to overcomplicate everything they touch...
Frightening to see "my software factory" upfront. Not "INSERT_STARTUP_NAME software factory", not "our factory", but "MY software factory". "MY".
If one person startup can "do" so much, how does it bode for broader software industry?
For as long as software engineering has existed, there have been prolific developers and there have been slow developers and every developer in between. Success has never been a product of volume. Some guy producing 30 "complex open source libraries" by hand in 2020 or with a "software factory" in 2026 doesn't mean anything. @sindresorhus published >1k packages long before code could be generated. Software is a means to an end, generating 30 "complex open source libraries" for no end changes nothing. If anything, these software factories are showing just how worthless most code is. Certainly, a "software factory" can be an accelerant towards building software that achieves some end, but it cannot magic an end out of nothing, it cannot make software valuable or useful, purpose must come from people.
> @sindresorhus published >1k packages long before code could be generated.
Interesting that you mention him, I was just thinking of him when I started reading your comment. Back in the day, he was famous for publishing lots of single-function npm packages like is-ip, is-regexp, copy-file, memoize, etc. Interestingly, he has now pivoted to mass-producing single-function macOS and iOS apps using AI. I guess some things never change.
Distribution has always been the bottleneck but that's becoming even more true now.
I have a one-person product and I'm exploring how to out-distribute incumbents who already have reach. I don't know if speed to market is a differentiator but cost surely is. Curious how others might be thinking about this.
>> I have a one-person product and I'm exploring how to out-distribute incumbents who already have reach. I don't know if speed to market is a differentiator but cost surely is. Curious how others might be thinking about this.
I have a one-person product that AFICT has no incumbents. It's an old idea of mine that AI could turn into an app in a weekend. Problem is if I finally make this, how do I get it out there before the clones come? I assume there are multiple players scanning the app stores for new things to clone and put under their already popular brand. One could get scooped even with a new idea. Am I off base with that concern? How to launch?
It's only a problem if the clones have solved distribution - which I don't believe they would have. Else they would have created a product the hard way without AI already.
I think if you can figure out how to find your customers and put your product in their hands, then AI is a true enabler. If you can't, AI doesn't help at all.
It’s not that frightening. Paying for tokens to keep this up and make sure it’s patched is boring work that folks would gladly outsource.
Also not sure why I’d trust one dude to be able to effectively manage this much context, or find all the bugs and issues. Or even patch them. I certainly wouldn’t trust myself to do it - and I’m an experienced engineer also doing a lot of work with AI.
Also don’t forget the luxury that is blaming other people for outages. It’s worth a lot of money.
Holy throwback, the author shaped my childhood apparently:
> Kurtis Welch has spent twenty years building and running production software.
> MCPVP A multiplayer game platform he started alone and grew to millions of players. Sold in 2015.
Good to know he's doing well. I'd buy him a coffee if I could.
Hahaha, hi. Small world.
will any of them be used?
Why?
[dead]
> they’re nearly 1.9 million lines of handwritten library code.
I don't know how innovative these libraries are, but this is surely an innovative use of the word "handwritten."
The line must come from the LLM, from that perspective, it is correct and virtually "handwritten" in spirit.
Also, 1.9 million lines of code isn't something to brag about, it's something to be ashamed of.
>> Also, 1.9 million lines of code isn't something to brag about, it's something to be ashamed of.
Seems like something the agents could try to fix. With all those automated tests, set them to work on reducing executable size (not code size).
lol makes me think of the invisible hand:
The invisible hand is a metaphor inspired by the Scottish economist and moral philosopher Adam Smith that describes the incentives which free markets sometimes create for self-interested people to accidentally act in the public interest, even when this is not something they intend.
https://en.wikipedia.org/wiki/Invisible_hand
> Recursive AI is Kurtis Welch and the fleet of AI agents he directs. You work with one senior engineer who owns the result, and you get the output of a full team.
Behold, a new type of consultant is born.
Thanks for the heads up, I'll avoid all of these libraries.
Did we forget that code is a liability?
And yet it's an asset at the same time.
Llms are great at cloning things they already exist. Is there anything novel in this list?
None are dumb clones. The ones that are the closest to "clones" are faster than competing libraries and have more developer friendly APIs and don't have external dependencies.
OneGame is novel.
One libraries maybe, zero products for sure
Interestingly, these libraries aren't the full factory output. It's just the part that's easiest to share. I can't exactly share production client code.
Congratulations?
Is it open source in spirit if the documents are slop and noone will maintain it?
This is the Nth software factory I've seen in 24hr. What now?
When did "complex" become something to boast about?
Hard things are complex. You want libraries to do complex things underneath simple APIs.
If libraries aren't doing complex things, you don't need a library to do it.
good for you
In 30 days of shipmas, this org has recreated libraries that already exist. Also, if you're going to just recreate open-source projects, shipping 1 a day is not impressive at all.
Anyone can rip off 7zip, imagemagick, lucene, spacy, whisper etc and call it innovation
>Full list to save people a click
>OnePdf — Read, repair, edit, sign and fill PDFs, and write PDF/A and accessible PDF/UA.
>OneDocument — Read, edit, recalculate and write Word, Excel and PowerPoint files, and render them to PDF.
>OneHtml — Parse and edit HTML like a browser, and check accessibility without one.
>OneRender — Lay out HTML and CSS and render them to pixels or tagged PDF without a browser.
>OneFont — Load fonts and shape text in any script, with bidirectional layout and line breaking.
>OneImage — Decode, edit, convert and resize images, from PNG and JPEG to layered PSD.
>OneLite — Embed a SQLite-compatible database that reads and writes the same files as sqlite3.
>OneIndex (preview) — Run full-text, vector and hybrid search in-process or in the browser.
>OneArchive — Compress and unpack ZIP, 7z, tar and other archives, with identical output across platforms.
>OneMail — Read, write and convert mail stores, including Outlook PST, and send and fetch mail.
>OneCrypto (preview) — Load keys and certificates, validate certificate paths, and sign and encrypt with OpenPGP.
>oneyaml — Load, dump and edit YAML in place in Rust, Go and C#, preserving comments.
>OneContact — Find, parse and normalize phone numbers, emails, URLs, addresses and names.
>OneDiff — Diff and merge text like Git, and redline changes between digital or scanned PDFs.
>OneModel — Run and train ONNX models in pure C# on CPU, GPU or in the browser.
>OneNlp — Tokenize, tag, parse and extract entities from text, and train models in C#.
>OneSpeech — Transcribe speech, add timestamps and identify speakers with Whisper and other open models in .NET.
>OneOcr — Read text, tables and invoice fields from scans and PDFs.
>OneLlm — Call Anthropic, OpenAI and Gemini models through one API and run agents defined in files.
>OneModelContext — Serve tools to AI agents over MCP or connect to MCP servers with OAuth.
>OneEcma — Run JavaScript inside .NET with execution limits and clear error reports.
>OneShader — Compile HLSL to Vulkan, Metal and WebGPU shaders as part of dotnet build.
>OneGpu — Run compute and ML kernels on Vulkan, Metal or WebGPU from pure C#.
>OneAudio — Decode, encode, mix and resample audio, measure loudness and play audio on devices.
>OneMedia — Remux, trim, tag, repair and package video and audio without re-encoding.
>OneGame — Let AI agents build, run, test, inspect and change games through commands without an editor.
>OneMath — Compute sin, exp, pow and more with correct rounding and bit-identical results across machines.
>OnePath — Find paths on grids, graphs and navmeshes, build navmeshes and steer crowds.
>OnePhysics — Simulate 2D and 3D rigid bodies, cloth, soft bodies, fluids and vehicles with bit-identical results across machines.
>OneScope — Scrub, fork and compare live or recorded runs of deterministic simulations.
Are any of these value add? I can create OneDiff in one line of code in .bashrc `alias OneDiff=diff`
Hilariously, none of these say what they do differently. OneModelContext, how is that different than MCP? It's description is literally a description of MCP, but it says it connects to MCPs. So what does it actually do? Reading the repo, it says it is an SDK for MCP, but then it also says it's an API for MCP. Which is it?
What value are these things if you can't even get the readme to agree with what it is? Why should I trust any of it to work correctly?
Additionally, the oneyaml's github link goes to this company's github profile, not a repo and clicking the link for oneyaml in the profile sends you back to this website. If you can't even link to the repo correctly, what are the odds of anything working correctly?
And for this, whatever (this) is people want to pollute our air with datacenter smokestacks.
Im not an environmentalist, but it seems wrong. 10,000 monkeys typing Shakespeare,
...NOW with Pollution!