I Built Smol Coder, the Best Free Claude Code

Author: Leon van Zyl · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I Built Smol Coder, the Best Free Claude Code · Channel: Leon van Zyl · 14:26 · Watch on YouTube

Topics: AI agents, Claude Code, Software engineering with AI, Local and open-source models, Open source and software ecosystems

Summary

Leon van Zyl introduces Smol Coder, an open-source coding agent designed to work efficiently with free local models. He demonstrates installing it, connecting it to Ollama or LM Studio, and using its terminal and web interfaces, including model selection, permissions, and an integrated browser. The video shows Smol Coder building a simple tic-tac-toe game and a more complex Minecraft-style project from a specification drafted with a more capable model. The creator highlights its planning and compaction features, uninterrupted work, and iterative debugging with screenshots.

Key points

Smol Coder is presented as a free, open-source coding agent optimized for local open-weight models.
The agent automatically detects models configured through Ollama or LM Studio.
The video demonstrates installing Smol Coder with npm and downloading a model through Ollama.
Users can operate Smol Coder in a terminal UI or a web-based interface.
Permission modes include read-only, edit, and bypass permissions.
Smol Coder uses a to-do plan to help the agent stay on track when context compaction occurs.
The integrated browser can display a project while its development server is running.
For complex projects, the creator recommends using a more capable model to draft a clear, concise specification for the smaller local model to implement.
Screenshots can be attached to help a vision-capable model identify and fix visual bugs.

Chapters

Start	Chapter	Summary
00:00	Why build Smol Coder	The creator explains why common coding harnesses can be inefficient for local models and introduces Smol Coder as a free, locally focused alternative.
02:34	Install Smol Coder and set up a local model	The video walks through installing Smol Coder, downloading a model with Ollama, and trying it in the terminal.
04:06	Terminal and web interfaces	The creator demonstrates model selection, effort levels, permission modes, project sessions, and context-window information.
06:40	Build and preview a simple game	After a sponsor segment, the creator has Smol Coder build a tic-tac-toe game and previews it using the integrated browser.
10:17	Plan and build a more complex project	The creator recommends preparing a concise specification with a more capable model, then has Smol Coder implement a Minecraft-style game.
12:53	Fix visual bugs and iterate	The creator shares a screenshot of missing textures with the agent, which fixes the issue, and concludes with advice to keep iterating.

Quotes

It’s open source, it’s totally free to use, and it’s highly optimized for free local models that run on your own hardware. — 01:33

This should detect the models irrespective of whether they’re running in a Docker container or just running on your machine. — 04:36

This plan actually keeps the agent on track. — 08:43

what I recommend is using a more intelligent model to create that initial plan. — 10:48

Smol Coder will ensure that this agent keeps working until it’s done. — 12:22

Concepts and entities

Smol Coder, Ollama, LM Studio, Qwen3.8, Claude Code, Oracle Agent Studio for Fusion Applications, ChatGPT, Three.js

Full transcript (original language)

00:00 I just created the best coding agent for free local models. I’m sure most of you are tired of watching agents build Minecraft clones, but this is different. This game, along with all of the animals and their artificial intelligence, everything that you see here was built by a free local model running on my own machine, and you can do this too. This means you don’t have to pay for API costs, your machine can be offline, and everything is running securely and locally on your own hardware. I’ve created tons of videos on my channel

00:30 showing how you can use free local models in coding agents like Claude, the Pi Agent SDK, DeepSeek Harness, OpenCode, and a lot, lot more. Now, as good as these coding harnesses are, they’ve got one massive problem. They’re not optimized for free local models. Open-weight models like Qwen 3.8 have very small context windows. This only goes up to 256,000 tokens if your hardware allows for that. The problem with running those models in something like Claude Code is Claude already takes up 34,000

01:02 tokens before you’ve done anything. And if your entire context window size is only like 128,000 tokens, this becomes a massive issue. If you want to load these models into any of these harnesses, it involves a lot of technical setup. I simply wanted a solution where I can start a new session, click on the model, it should automatically detect all of the models I have set up on my machine whether I’m using Ollama or LM Studio. No configuration, no setup, it just works.

01:33 So that is why I created Smol Coder. It’s open source, it’s totally free to use, and it’s highly optimized for free local models that run on your own hardware. I’ve pretty much optimized everything for local models from the system prompt that gets injected at the start of the conversation, the tools that are available to the agent, the way compaction works, and we’ve introduced a special to-do list that keeps the agent on track after compaction happens. If you would like to find out more, I’ll link to the GitHub repository in the description. And while you’re

02:03 there, be kind and hit the star button. In this video, we’ll go over setting up Smol Coder, we’ll also download our very first open-weight model and run it on our own machine. And finally, I’ll give you a really cool tip for building complex projects like this Minecraft clone. Also, if you find these kinds of videos useful, hit the like button and subscribe to my channel for more agentic coding tutorials. Alright, to get started, just open up the command prompt or terminal, whatever you want to use. PowerShell is perfectly fine as well. Then to install Smol Coder,

02:34 we’ll just run this command npm install -g smolcoder. I also like to just tag @latest to ensure that we always install the latest version. And done. We can now start using Smol Coder, it was really that easy. Before we start Smol Coder, we do have to download a local model onto our machine. For this, we basically have two options. We can use Ollama, which is incredibly popular or LM Studio. For this tutorial, I’ll go with Ollama. And if

03:04 you want to follow along, go to ollama.com, then click on download, select your operating system, and then download and install Ollama. Once you’ve installed Ollama, we can browse all the different models and there are tons of models available. The choice of model will depend on your hardware. I think Qwen 3.8 is the best free model for coding tasks. Now the model itself is about 18 gig, starts about 16 gig of VRAM or unified memory if you are using Mac. To download the model, we can simply copy this

03:35 command, then back in our command prompt or terminal, we can run the command Ollama pull, followed by the model name. We can then check if the model works by running Ollama run, then we can just enter that model name again. The first time you run this, it will take a while as the model loads into memory. Afterwards, we can send it a message like, “Hey,” and we get the response back so we can see the reasoning effort as well as the agent’s response. We can use Smol Coder in one of

04:06 two ways. We can run it in the terminal or in the web-based view. If you want to run it in the terminal, I do recommend opening up Smol Coder in a very specific project folder. For example, create a new folder on my machine and I’ll just open the terminal in that specific folder. Then we can run the command smol. This will open up Smol in a terminal UI. From here, we can select the model by running front slash models. When we press enter, this will automatically detect the

04:36 models running on LM Studio and Ollama. This should detect the models irrespective of whether they’re running in a Docker container or just running on your machine. Now we can just select the model that we would like to use. We can also change the effort levels by running the command front slash effort and I’ll just select high effort. And of course, if we send a message, we can see the response streaming in immediately. We can also change the permissions mode by holding shift and pressing tab. So we can go from edit to bypass permissions to

05:07 read-only mode. In read-only mode, the agent can’t make any changes. It can only read the files in the project folder. In edit mode, it can edit the files, but it can’t run commands without asking your permission to do so. In bypass permissions mode, the agent will be able to change files and run commands. We can also exit out of this project by running the exit command. Now I know a lot of you don’t want to use the terminal view and you prefer the web-based version. In order to start the web-based version, just run the command smol dash dash web.

05:38 This will give you this unique URL. What’s cool about this is you can see all your different projects on the left sidebar. So if you’ve ever used Claude Code or Claude Desktop or the ChatGPT app, it’s very similar. So you can view all your different projects and sessions over here. Let’s start by opening a new folder. So I’ll just select that smol demo folder and I’ll enable start a session. Cool. We can see our new project over here and our new session. As with the terminal view, we can also change our permissions mode. So I’ll just leave it on edit. We

06:09 can also change the model, which will automatically detect the local models on your machine. We can also change the reasoning effort. I’ll leave it on high from here. We can also see the context window at the moment. I’m using about 1% of the context window and we can see the context window size is a hundred thirty one thousand tokens and not the full 256,000 tokens. That is because Ollama will automatically look at my VRAM usage on my machine and dynamically allocate the context window that

06:40 is available to this agent. Also keep in mind, I am currently running my recording software, which is also taking up a lot of my VRAM. So let’s just say, Hey, and we can see the reasoning output over here, as well as our agent’s response. Let’s build our very first project. We’ll start with an easy one and then afterwards I’ll show you a more complex project and how you can get the best results out of these tiny models. But first, a quick word from our sponsor Oracle. If you’ve ever built an AI agent that works great in a demo, you

07:10 know what happens next. The moment a real business needs to use it, everything gets hard. Who can access what, who approves changes, who keeps track of it all. And usually that’s where these projects get stuck. Oracle Agent Studio for Fusion Applications in a smart way. Your agents run inside Oracle Fusion itself. So all the security rules and tracking that companies already trust, your agent gets that for free. And here’s the part I like. You build with the tools you already use

07:41 every day. That could be VS Code, Cursor, the terminal, Git and coding agents like Codex and Claude Code. There’s also a public GitHub repo full of templates, and sample apps. So you’re never starting from scratch. Simply go to GitHub.com slash Oracle slash Fusion AI Studio. If you’re already a Fusion customer, AI Agent Studio is included at no extra cost. And the GitHub repo is open to everyone. Go to Oracle.com slash

08:13 applications slash Fusion AI to check it out. The link is in the description. Thanks again to Oracle for sponsoring this video. For our first project, let’s keep it light and simple. I’ll give a simple instruction like create a tic-tac-toe multiplayer game that runs in a browser. Now some of these models do support vision as well. So if you do want to attach images as reference, you can do that as well. Either by clicking on this attachment button or just pasting from your clipboard. I’m just going to send this. Now this is one of the

08:43 core features of Smol. You see this plan over here. This is slightly different to the plan that you would see in something like Claude Code. This plan actually keeps the agent on track. So if compaction had to happen, and because this model’s context window is so small, compaction will happen often. And the agent will then review this plan to figure out what was built and what it needs to do next. And as it completes all of these tasks, it will actually go and tick off all of these plans. And that was really, really quick. And then the game

09:15 was built. But how do we now actually view this game? Well, we could just open the folder and run the file from here, or we can get the agent to start a dev server. Smol does include an integrated browser. So if we want, we can just view the game in this little browser without having to leave this view. So we can just say, please start the dev server. Because we’re in edit mode, it is going to ask our permission to run this command. And done, right. So it’s running the game on this URL,

09:45 which I can simply paste into the integrated browser. And this port is actually still running that Minecraft clone from earlier. So I’m just going to let it know that that port is already in use. And now the agent is running our game on a different port. So and there we go. So we can see the game over here. And if we take this little button, it will open the website in its own little browser. And there we go. Now we’ve got a fully functioning tic-tac-toe game. And this was built in no time. This is so freaking optimized. Alright,

10:17 so for the next game, I am just going to delete all the files in this folder. I’m going to start a new session. And let me just show you how to build complex games and projects like this Minecraft game. Now the general rule of thumb is that these tiny models might be good at sort of implementing changes. But that initial plan and the initial set of requirements need to be extremely clear. We have to be honest with ourselves that these small agents will never be at the level of the flagship models like ChatGPT or the Claude models. So what

10:48 I recommend is using a more intelligent model to create that initial plan. And then we can give that plan to our small model to implement. And yes, this can be totally free as well. In fact, I’m not even signed into ChatGPT. And what we can do is this, this can give me a spec that I can create a smaller free open source model to build a Minecraft clone. We can keep it simple, keep it only a few biomes. But give it enough instructions to create a game using HTML, CSS and JavaScript, it should be able to run in the browser. You can use things like Three.js, the

11:20 textures, the skybox and the sun should look accurate. There should be a few animals as well. And the animals should have some very basic AI. There should also be a very simple crafting system. These models have limited context windows. So just be very careful with your wording. Don’t make it too long. Let’s send this. Then what we can do is take the spec generated by ChatGPT, simply paste it into Smol. So this is the complete spec. And we can go ahead and send this. And we can see this agent is thinking very hard. We can expand this

11:52 thinking line. So we can actually view all the reasoning at the moment. We can also easily collapse this if we don’t want to see it. And of course, we’ve got our plan showing up and we can see exactly what the agent is working on. Apologies if the webcam is lagging. This is taking up a lot of VRAM on my machine at the moment. So I’ll let the agent complete and then we’ll have a look at the final result. And we’re done. So it took about 20 minutes. And if you’ve ever used these local models and other harnesses, you know that these

12:22 tend to crash and you kind of have to wake them up and ask them to continue. Not with Smol Coder. Smol Coder will ensure that this agent keeps working until it’s done. So this ran for 20 minutes. It was uninterrupted. I never had to ping it or restart it. So I just asked the agent to run the game on port 1234. So let’s grab this URL. Let’s pull up a browser. And this is Voxel Mini. What I can see, in fact, let me just open this up in full screen mode. What I can see is that some of these textures are messed

12:53 up. So if we kind of look at the world, we can see that some textures are actually missing. So what we can do, we can actually screenshot this. So I’m just going to grab a screenshot of this bug. I’ll just do this. All right. Then back in Smol, we can actually just paste in that screenshot or attach it. And we can say, hey, it seems like there are textures missing. Right. Let’s send this. So vision needs to be supported by the model, by the way. So

13:25 Qwen 3.8, it does support vision, tool calling, thinking. So whatever model you’re using, just ensure that it’s got a tools and vision tag as well. And let’s have a look. And by the way, this is normal with these small models. You kind of have to nudge them a bit. So you would take screenshots and just let them know that something is broken. But eventually you’ll get there. And it’s actually a very good way to kind of learn how to prompt agents effectively. Right. So after refreshing this game, let’s just see if this bug is fixed. And yes, it is. So

13:55 just look at that. Now we can actually see the textures. And this is cool. You know, we’ve got our little game. Everything seems to be working. We’ve got animals. And yeah. So, you know, we can keep reprompting it. We can keep iterating on it. We could even ask, you know, ChatGPT or Claude just to come up with a more detailed plan for us using their free modes. And yes. So let me know what you guys think. You know, definitely try out Smol Coder. Let me know down in the comments what you think about these types of harnesses. And again, hit the like

14:26 button and subscribe to my channel if you want to see more agentic coding tutorials. And I’ll see you in the next one. Bye bye.