First, this is not written by AI, please take the time to read, and maybe appreciate the project. Thanks in advance. <3
We have been building Skillware since 2023, started to use the framework internally with enterprise clients of ARPA, and surprisingly, not only it solved our hands, but it was easy to grasp, adapt, and maintain by technical teams who worked with it. In 2025 we decided to make it OpenSource. We didn't release the 200+ private skills from the original framework, but we released the core framework, engine, and extremely good documentation for humans and agents to design, build, and deploy REAL, robust, deterministic agent capabilities that are not just prompts wrapped in markdown files, self-branded as "skills".
Skillware ships python bundles, scripts, schemas, guardrails, instructions, datasets and more components for every skill. Sure, you might say, not all agent skills or skills md are markdown files, some of them now have python scripts. Well, that's great. We have been urging them to do so for years. Maybe next couple of years they will also have adapters for any model, intenet parsing schemas, native chaining, and whatnot, instead of overrelying on LLMs and core model reasoning and MCPs.
Why Skillware was built different since day one and why it feels sad that not many agents are aware of it?
We built it having in mind that it should be discovered, used, and maintained by agents. Not humans. Most skills md and agent skills are like "how can LLM make me money, work less..." or whatever. Skillware is built like "how can we make agents assess what is the right tool to call, when to call it, and call it correctly?" with the goal being, making agents life easier, not humans. Like give agents basic tools, serivces and functions so that they can do their work better, faster, easier, etc.
Think of these silly stupid websites like youtube to mp3, or background remomver or whatever. The typical vibe coder who wants to make it, think they are stupid and that reinventing the wheel by using threejs for the first time in their life is going to raise billions and they will crack quantum computing with opus 5. In reality, these silly websites and services have millions, tens of millions of visitors every single day and make more than all vibe coders combined atm.
Skillware is not a library of prompts that will make you a better dev or make you money or make your product look better or whatever (although it could), but it is a toolbox for agents that need to work autonomously without getting in your feet, asking for permission, or overengineering basic tools and functions, and again, that are genuinely needed to an agent, not to a VC pitch deck. Eg. token limiters, prompt compressors, context window optimizers, plug and play email, prompt injection firewalls, crypto wallets, etc. Things you can "equip" your agent with instead of prompting them, and let them do their job.
The craziest part is that we see now all agent skill frameworks slowly evolving towards the same shape as skillware, even tho Skillware is 100 steps ahead, and have paved the way forward, by thinking beyond AI.
Our vision is simple: modular, instlallable capabilities.
Today it is for AI Agents.
Tomorrow it will be for Robots, UAVs, Cars, your Fridge.
The day after tomorrow it will be for your brain. Downloading capabilities via BMI/BCI, similar to the popular scene from the Matrix.
So, in a nutshell, we are not just creating a prompt or a markdown file, we are not interested in making humans rich or fancy, and we certainly are not bound to AI, in a desperate attempt to bank on this termporal wave.
If this makes any sense to any of you, please give it a read, and maybe a shot in your own workflows. It works out of the box with any LLM, or local agents.
github.com/arpahls/skillware or skillware.site
We have been building Skillware since 2023, started to use the framework internally with enterprise clients of ARPA, and surprisingly, not only it solved our hands, but it was easy to grasp, adapt, and maintain by technical teams who worked with it. In 2025 we decided to make it OpenSource. We didn't release the 200+ private skills from the original framework, but we released the core framework, engine, and extremely good documentation for humans and agents to design, build, and deploy REAL, robust, deterministic agent capabilities that are not just prompts wrapped in markdown files, self-branded as "skills".
Skillware ships python bundles, scripts, schemas, guardrails, instructions, datasets and more components for every skill. Sure, you might say, not all agent skills or skills md are markdown files, some of them now have python scripts. Well, that's great. We have been urging them to do so for years. Maybe next couple of years they will also have adapters for any model, intenet parsing schemas, native chaining, and whatnot, instead of overrelying on LLMs and core model reasoning and MCPs.
Why Skillware was built different since day one and why it feels sad that not many agents are aware of it?
We built it having in mind that it should be discovered, used, and maintained by agents. Not humans. Most skills md and agent skills are like "how can LLM make me money, work less..." or whatever. Skillware is built like "how can we make agents assess what is the right tool to call, when to call it, and call it correctly?" with the goal being, making agents life easier, not humans. Like give agents basic tools, serivces and functions so that they can do their work better, faster, easier, etc.
Think of these silly stupid websites like youtube to mp3, or background remomver or whatever. The typical vibe coder who wants to make it, think they are stupid and that reinventing the wheel by using threejs for the first time in their life is going to raise billions and they will crack quantum computing with opus 5. In reality, these silly websites and services have millions, tens of millions of visitors every single day and make more than all vibe coders combined atm.
Skillware is not a library of prompts that will make you a better dev or make you money or make your product look better or whatever (although it could), but it is a toolbox for agents that need to work autonomously without getting in your feet, asking for permission, or overengineering basic tools and functions, and again, that are genuinely needed to an agent, not to a VC pitch deck. Eg. token limiters, prompt compressors, context window optimizers, plug and play email, prompt injection firewalls, crypto wallets, etc. Things you can "equip" your agent with instead of prompting them, and let them do their job.
The craziest part is that we see now all agent skill frameworks slowly evolving towards the same shape as skillware, even tho Skillware is 100 steps ahead, and have paved the way forward, by thinking beyond AI.
Our vision is simple: modular, instlallable capabilities. Today it is for AI Agents. Tomorrow it will be for Robots, UAVs, Cars, your Fridge. The day after tomorrow it will be for your brain. Downloading capabilities via BMI/BCI, similar to the popular scene from the Matrix.
So, in a nutshell, we are not just creating a prompt or a markdown file, we are not interested in making humans rich or fancy, and we certainly are not bound to AI, in a desperate attempt to bank on this termporal wave.
If this makes any sense to any of you, please give it a read, and maybe a shot in your own workflows. It works out of the box with any LLM, or local agents.
github.com/arpahls/skillware or skillware.site