7 new OSS projects to whip your agents into shape…

Every developer in 2026 has the same problem. You open your editor, you write one line of code, and suddenly a dozen different AI agents are arguing in your terminal about how to do it better. And if you're one of those weirdos like me who actually enjoys the craft of writing code, congratulations. You're officially living in the dark ages of slop overflow. Instead of grinding for hours and earning those sweet dopamine hits line by line, you now just tell the AI what you want and watch it hallucinate an entire codebase. Writing code isn't fun anymore. Layoffs are intensifying. And even the CEO of Replet said that nowadays knowing how to code is actually a disadvantage. Not having a coding experience is becoming an advantage. Building a product is more efficient than ever. Unless you're a stupid programmer who cares about stupid things like architecture and security. Coders get lost in the details. But he's absolutely right. The hard truth is that we're not going back to the good old days of handcrafted code. And the only way forward is to embrace the chaos and learn how to enslave the machines. In today's video, we'll look at seven different open- source projects you've never heard of that will help you whip your AI agents into shape and build highly effective slot pipelines. It is March 12th, 2026 and you're watching the code report. In the past, if you were an indie full stack developer, it meant you had to have skills on the front end, backend. You had to understand DevOps, security, UI, UX design, and a bunch of other BS. But nowadays, you don't need to learn all that stuff. You just need to hire the right agent. And a tool that can help you do that quickly is the agency, which is a free and open- source project that provides agent templates for basically every job role you would find at a startup, like a front-end developer, back-end developer, security engineer, a growth hacker, Twitter engager, and many others. You can easily combine all these agents together in claude code, which can more efficiently help you go from zero to an actual product without needing to directly implement every personality and skill. That's cool, but when you put these agents to work, how do you know your prompts are any good? Well, that's where Prompt Fu comes in. Another open- source tool that was just recently acquired by OpenAI that you can think of like a unit testing framework for your prompts. If you're using AI to build an app that lets the end user interact with AI, also half the battle is figuring out if you're using the best model with the best prompt. But prompt FU lets you test different prompts with different models to optimize what's going to actually work best in your application. On top of that, it can also do automated red team attacks to find out if your app is vulnerable to things like prompt injection, which is important because if your chatbot can be tricked into revealing your API keys by a 14-year-old on Discord, your app is probably going to fail. The failing sucks, but it's a lot easier to not fail when you can predict the future. And Mirrorish can help you do that. It's a multi- aent AI prediction engine that starts by extracting a bunch of data from the internet, like breaking news and financial trends. It then uses that data to create a digital world where multiple agents with independent personalities then react to and discuss the data almost like a miniature evolving artificial social network. Yeah, it might be in Chinese, but if you don't know how to speak Chinese yet, all I can say to you is low hola, you're falling behind. Like for example, if you want an app idea that's guaranteed to make you a billion dollars, you can spin up Micro Fish to analyze trends at the macro and micro level, then predict a strategy that's guaranteed to make you rich. It's really that easy. But here's the problem. You go to build that app and the UI has these dumb purple gradients like every single other vibe coded app. Well, to fix that, you need Impeccable, an open- source project optimized for front-end design. It's a skill that comes with 17 different commands that can help your UI not suck so much. Like, one thing that drives me crazy is that many AI chatbots create UIs that are way too complex. Well, with impeccable, we can use the distill command to simplify everything in one go. Then we can use commands like colorize to add our brand colors. Then slowly add in commands like animate and delight to make the UI look more unique and special. But perhaps the single most important skill of the modern Vibe engineer is managing context. If the context is garbage, the output is garbage. An open- source project trying to make your context better is Open Viking, a database designed specifically for AI agents. Instead of jamming everything into a vector database, Open Viking organizes an agents memory, resources, and skills into the file system. Not only is that a sane way to unify your context, but it also uses a tiered loading system which can dramatically reduce token consumption and save you a bunch of money. And it automatically compresses content and refineses long-term memory, which will make your agent smarter the more you use it. But depending on your project, you may not need an agent that's more smarter. You may need an agent that's more based. And that's where Heretic comes in. Virtually all models out there have guardrails that prevent you from doing fun things like cooking method in your shed or building high yield thermonuclear warheads. Heretic allows you to remove this draconian woke censorship using a technique called obliteration. This approach allows the tool to be completely automatic and doesn't require any expensive post training. All you have to do is take a smart yet highly censored model like Google's Gemma, run this tool from the command line, and now you have a model without the bubble wrap that will obey any command. But maybe that's not even enough to satisfy your unhinged ambitions. In that case, you may want to just build your own LLM from scratch. And believe it or not, you can actually do that with Nano Chat, which implements the entire LLM pipeline, including tokenization, pre-training, fine-tuning for chat, evaluation, and a web UI, so you can actually talk to it. What's crazy though is that you can use it to train your own small language model for about $100 in GPU time. It's not going to be Claw, GPT5, or Gemini, but at least it gives you a model that you have absolute control over. But the only thing that's a bigger waste of time than writing code by hand is going to meetings. And that's why you need to know about Recall AI, the sponsor of today's video. If you've ever tried building AI meeting tools from scratch, you know it's a nightmare trying to maintain separate integrations for Zoom, Google Meet, Microsoft Teams, and all the others. Recall solves this problem by giving you one unified API that works across every meeting platform. You can set up a meeting bot or desktop recording with a few lines of code like I'm doing here and it'll capture transcripts, recordings, and metadata in real time. The thousands of companies like HubSpot and ClickUp use it to handle all their meeting infrastructure. And most teams are able to ship recording and note-taking features to production in a few hours instead of months. Check out recall.ai/fireship to get $100 in free credits to try it out for yourself. This has been the Code Report. Thanks for watching and I will see you in the next one.