Unconciously I find myself trying out a new AI/LLM tool every Monday. There's something about spending time with the family in the weekend that makes the mind realise what experiments need to happen and these ideas typically lend themselves very well for new tools.

This week I tried out cline through their VSCode extension. These are my notes after kicking the tires.

Plan vs. Act

This might be the main feature of Cline that made is stand out for me.

Cline gives you a conscious "plan" and "act" mode. This sounds like an odd feature at first, but it's a great forcing function to make sure the agent has enough context. Cline won't start writing code unless you trigger it to go into "act" mode. I almost wrote off cline at the start because it starting doing all sorts of silly things but then I realized this feature was there. Once I found it, smooth sailing all the way. The keyboard shortcut for switching between these two is cmd + shift + A, but you may want to change this shortcut depending on other extensions (it's a popular shortcut).

Important!

If you're used to claude you might be tempted to think in "act"-mode but if you want to benefit from cline you really need to let that go. When you spot a bug, context really matters, so you want to make the agent go into "plan" mode first. Not always, sure, but 90% of the time it saves a lot of time and tokens.

Tokens

Cline is not making money off your tokens usage which is reflected in a bunch of transparancy.

You can clearly see the distinction between prompt/completion tokens as well as the token cache read/writes. You can even point to your local installation of claude and it will know how to work with it. It's also possible to use models through their API and they seem to offer a good selection to pick from.

Task done

Another sublte little UI detail is how the text turns green when it is done with a task.

When cline thinks it is done with a task the text will turn green. It's a tiny detail but it does activate something in the brain that makes you switch modes too. Note that when you're not pleased with the final result of a task that it is pretty easy to go back and to restore the state at a previoius point of the conversation. This was a great feature from Windsurf and I am very happy to have a way that let's me add this behavior on top of claude code.

Small gripes

I honestly feel like there is a lot to like here for personal use but there are a few small gripes, which are probably just bugs that will get fixed down the line.

Cline might run a bunch of terminal commands for you and it will start up a bunch of terminals for them. It doesn't seem like it can close them, although this also just be related to me switching a few models.

The other gripes I have are more related to shortcuts. I can auto-complete mouse clicks thanks to homerow but I would prefer that I don't need to do this. The "accept"/"deny" buttons don't show the relevant shortcut on hover, which feels like a bad oversight.

The main feature I am hoping for is that they make it easy to add multiple models for different modes. As this blogpost mentions: it might be a honking great idea to use Gemini for planning (with it's huge context to check all files) and to use another model for coding (Kimi K2 maybe?).

All these things can be fixed, and they probably will, but worth pointing out that there are some bits that don't feel 100% done just yet.

Verdict

I prefer cline now as a "sidebar" experience. You can also use this extension together with Cursor so if you want to keep using that for tab-completion, so this might be my IDE tooling for this week.