How Cline Quietly Changed The Game For Code Copilots

This is a short essay about how Cline marks a shift in how we use LLMs compared to tools like Cursor.

Working in San Francisco and having seen Cursor's rapid early word-of-mouth spread, I'm constantly surprised how Cline, which seems an order of magnitude better, has not had the same viral word on the street. Better products on paper often lose to their competitors, but the key thing that seems overlooked is how the AI code-generation landscape is changing right under our feet.

I heard of Cline through one of my teammates (an absolute stallion) who lives in Ukraine. Skeptical at first, I tested Cline against Cursor for the first few code changes and struggled to see what this genius teammate saw in Cline. At some point, I hit that wall that you often hit with Cursor, where you need to roll up your sleeves and do things yourself. But Cline pushed through the task in our project with 60k lines of code without skipping a beat. Cline kept succeeding for the next few complex code changes I tested, whereas Cursor failed. Ever since then, Cline has been my go-to copilot for coding. (Fun fact: I used Cline to tell me how many lines of code we have.)

I’m not alone in this experience; some of my founder friends also didn’t get it immediately, but they all saw the magic after a bit of perseverance.

Surprisingly, many of our users at Coplay lean toward Cline and Roo Code instead of Cursor, probably because Cline offers an uncompromised free tier if you use your own API key. In this case, you only pay Anthropic (or similar) for what you use.

Cline Out-Competing Cursor

Surprisingly, Cline makes it look easy to out-compete Cursor regarding functionality. Before seeing Cline, I thought it would be impossible for most companies to compete with Cursor’s product because they have a small school of IMO medalists all building carefully calculated infrastructure to ensure you get fast and accurate context fed to the LLM.

But, after looking at Cline’s code, I saw the beauty in how such a simple wrapper could unleash new levels of productivity when combined with the more recent AI models such as Claude 3.7 and Gemini 2.5. In contrast, Cursor’s RAG and caching optimizations make working with older AI models feel like magic but restrict the more recent models too much.

I never thought we could compete against Cursor for code generation and editing for the same reasons mentioned earlier. But thanks to the shift brought about by tools like Cline, we can provide a code generation experience as good or even better than Cursor.

A Different Way of Using AI

Cline is almost a different product altogether. The typical workflow with Cursor is to describe your task to the agent, then look at the changes and accept or reject as you see fit. Thereafter, you re-prompt again to make more changes and repeat. With Cline, your workflow changes into more of a real-time chat with an assistant, as if you can steer the agent.

For example, I might ask Cline to refactor my database operations to be more performant. Cline will read the relevant files and start suggesting changes for my approval. If at any point, I don’t like the direction it’s taking, I can start steering the agent toward what I want early in the editing process instead of waiting until the end. Cline might decide to tweak my SQLite implementation, but I can intervene and steer it to focus only on Supabase interactions because I know that we’re deprecating SQLite-related code soon.

No More RAG

This workflow, introduced by Cline, effectively sidesteps RAG and lets the model decide what parts of the project to read to complete the task. The initial context provided in Cline is quite simple -- a breadth-first traversal of the files in the project up to 200 files. From there, the model can take over and traverse the project as needed.

This agentic project ingestion makes Cline more expensive than Cursor, and I often have single threads in Cline that cost over $6. But it's well worth it when it's $6 to solve something that Cursor cannot.

(Note: all of my comparisons to Cursor are with settings turned to using maximum context.)

All that said, Cursor still has a huge moat – the best tab-complete model on the market. The Babel model they acquired has a mind-blowing 1M context window and 250ms response time. With their tab-complete moat and the fact that you can install Cline inside Cursor, Cursor still has a favorable position in the market. However, my use of tab-completion has dramatically reduced after learning how to steer the Cline agent -- more on this in a few weeks.

As the pace of AI innovation continues increasing, I’m sure we’ll see paradigm shifts like this happen more often. I wonder what the next update will be to uproot the Cline way of interacting with AI.