Vibe coding with Cline
This is a screen-grab of the Cline panel whilst it is doing a task. As you can see it is very verbose. I guess it’s nice that you can see both costs and context window size in the Task bar.
Personal takeaways
- I like Cline a lot as an in-IDE assistant. I like that it integrates with OpenRouter and that allows me to code using lots of different models.
- It does seem unnecessarily verbose. It feels like the context for a simple request is probably a lot bigger than it needs to be.
- The benefits of coding in this way may be greater for a larger project with lots of files. However I can also see the costs growing very quickly for that. For the style of coding that I like—where I try to break down a larger task into a lot of smaller scripts that can be one-shotted and then refined—coding in this way may not be much more useful than just using the Chat interfaces with MCP to it can write files. I need to test that.
I’ve not done much experimenting with in-IDE vibe coding. Of the options available to do this I like Cline. It has all the main features I’m looking for I think:
- It’s Open Source.
- It works with Visual Studio Code.
- It can work with multiple LLM providers, including multi-LLM services like openrouter.ai.
- It’s active and maintained on GitHub, with currently 4.3k stars.
I am going to use it with openrouter.ai so that I’ll be able to test Cline with various models. I added 20USD of credit to openrouter.ai..
Configuring Cline is very easy, and there are a reassuring amount of “Are you sure you want to do this” type messages from both VS Code and openrouter.ai. I put a limit of 20USD on the key I gave Cline access to, just whilst I am testing.
So I did a little test, getting it to create a spinning “hello world”, which it did successfully, using Claude 3.7-sonnet. But then I asked it to make it expand and contract as well, which it didn’t succeed at, despite claiming that it did. Which is disappointing as this is a very simple test. But it’s not an issue with Cline.
Cline asks for confirmation for every step of the process, which is a lot of confirmation. It is configurable so that it only needs to ask for approval for certain actions, such as writing to the file system. I may just get it to approve everything once I’m used to it.
It constantly updates the credit figure. The above cost 27 cents. Which is nothing of course but I’m not sure I want to be constantly reminded that I am using credits, so I may turn that off. It does make me appreciate my ChatGPT account though, I use that constantly throughout the day and I’m sure I must use more that 20USD credit each month, in fact if I was billed for usage I’m sure the bill would be much higher.
So do I like coding like this? I’m undecided. It’s not quite as great as I imagined, it feels like you have less control over the process because Cline is deciding the steps to take. I guess I could try using much more specific prompts. Also I need to test coding in the Claude app itself using a MCP to allow it to read and write files locally. More testing needed.