Transcription
[Music] [Applause]
Hi, welcome to another video. So, ever since Claude Code basically started rug pulling, I have been trying to find some better alternatives to it. And today's video is about one such alternative called Codebuff. I actually saw it on Open Rotor's usage leaderboard and thought I'd give this a try.
It is a CLI tool and it is open-source as well, but you can't use any provider with it as of now. You can only use their own Codebuff provider that works on a credit system, which is basically just a wrapper over Open Router and allows you to use all of the Open Router models. Though you can't directly change the models, each agent that it has uses a different kind of model. Like for planning, it can use GPT-5, but most of the stuff mainly uses Sonnet. It gives you 500 free credits monthly, which basically translates to about $5 worth of credits. I would have liked it if it allowed us to use our own providers, but that isn't available yet as it's only recently been open-sourced. So, I hope someone adds that option to it, maybe makes a form or just makes a pull request.
Anyway, that's not too bad because the credit system is way better than the membership model that is not sustainable and just ends up in rug pulling. So this is fine. How it works is pretty interesting. It has some sub-agents that are specialized for specific tasks like testing, nitpicking, and stuff like that. It can spawn those agents on its own, and it actually works well. You can also make your own custom agents in the syntax they have provided as an example. You can also use the SDK they give to build out scripts and custom workflows. There's also an agent marketplace where you can add even more agents to it.
Now, I tested it with everything set as default, and I think it's great even in default configuration. That's what I used, and I've also tested it on the four questions that I usually use to test these tools. So, let's get into it and check it out.
But before proceeding, let me tell you about Ninja Chat. Ninja is an all-in-one AI platform where for just $11 per month, you get access to top AI models like GPT-4o, Claude 4 Sonnet, and Gemini 2.5 Pro, all in one place. I've been using Gemini for quick research. But what's really cool is their AI playground where you can compare responses from different models side by side. Their mind map generator is a game-changer for organizing complex ideas as well. The basic plan gives you 1,000 messages, 30 images, and five videos monthly with higher tiers available if you need more. Use my code KING25 for 25% off any plan or KING40 yearly for 40% off annual subscriptions. Check the link in the description to try it yourself.
Now, back to the video. You just install it with a simple command. Sign into your Codebuff account, and you're good to go. It's pretty simple and doesn't have any flair or polish, which is fine, and I won't judge. But that doesn't mean it doesn't have power either because it has got some good options. If I hit the slash command here, you can see some useful slash options. You get the `init` option that basically makes it index the project and generate some markdown files. `Diff` shows you the file changes done in each session. It also has checkpointing features, and you can revert back to any checkpoint or state. `Usage` shows you the credits consumed for each session and everything. `Reset conversations` will reset the conversation. `Compact` will summarize the conversation until now and basically compress your context size. There are also multiple modes: `ask`, `light`, `normal`, and `max`. And generally, I keep it on the `normal` mode.
Let's do something with it as well, which is also the first question that I usually test, which is asking it to build a movie tracker app using the TMDB API. And it goes ahead and starts to work. It doesn't ask for any approval, which is interesting. I'm fine with that, but I know a lot of people like to have approval options because that way you can stay in control. But it isn't available here. You can also see that it spins up sub-agents, plans, and executes everything as well. In a bit, it gets done. And if I show you what it made, then you could argue this is one of the best generations. It's arguably better than what Claude Code makes using the same Sonnet model, though Codebuff probably mixes in some other models, too. But even then, this is generally cheaper, and I'll talk about cost in a bit. Still, the result looks pretty awesome for sure. Codeex wasn't really good in this case, though. The GLM coding plan was pretty good and kind of cool for the value you get out of the $3 plan.
If we move to the second question, which is to make a graphical retro calculator in the terminal using Go and the Bubble Tea framework. This is a question that usually only GLM passes and no one else, but Codebuff also passes this, and it was one-shot. It works well, and you can see it has a good interface. I can navigate, and everything works properly, which is pretty cool. Claude Code's version doesn't work. Neither does Codeex, but GLM works. It's good, but arguably not as good as this. So, this is kind of awesome.
Then we got the Godo question where I give it a basic Godo FPS shooter template and ask it to implement a step calculator that tracks steps in a bar at the top, which is also configurable in the settings. I also asked it to implement a life bar that decreases after jumping, and it worked flawlessly in one-shot. You can see both the lifebar and step calculator at the top, and it's awesome. Claude Code also does this flawlessly, and so does Codebuff. So I like this generation a lot.
Then we come to the fourth question, which is the one that has been unbeaten, at least for CLI tools, and this is the first CLI agent that actually passes it. Here I give it the Open Code GitHub repo and ask it to add a new SVG generation command to it. The command should open an interactive modal where I can enter a file name and prompt, then make an SVG and save it. And it worked really well with this. You can see that when we use `/svg`, it gives us a modal in the same style as Open Code, where I can enter the prompt and file path, and it works perfectly. It's literally prompting Open Code itself, and it's great. It worked in one-shot, and it is awesome. I am really impressed by how well it has worked for me.
And it also seems that it's not intended to be a rug pull either because the pricing is transparent, and you only pay for the credits you use, which is good. I would still like if direct Open Router usage could also be enabled, but that may be a long shot because some agents cost per call as well. So yeah, I know my testing for agents is still rather limited, and I've been adding more new tests, but these are the main ones that matter for now. It leaps ahead of Claude Code by quite a lot. I like the code quality here a lot. It's way better than Codeex, Claude Code, and GLM.
It does come at a cost, literally. It can be costly at times. It's like two times the cost compared to Claude Code. This happens because some agents cost more credits and some cost less. Honestly, I think it would be much better to just have transparent cost directly instead of credits. Credits just feel less clear to me. It would be better to see it in terms of actual cost. I think it's a really good alternative to Claude Code. It's a bit costly, but the work it does is absolutely great. I've been using it, and it feels like chef's kiss. Really good stuff. There are some downsides, but the upsides make it super worth it. You can't even compare it to subscription platforms because those are designed to rug pull. So, there's that. Overall, it's pretty cool.
Anyway, share your thoughts below and subscribe to the channel. You can also donate via the Super Thanks option or join the channel as well and get some perks. I'll see you in the next video. Bye.
[Music]