Most AI coding tools give you suggestions. Cline does something different — it acts. Open a task, describe what you want built or fixed, and Cline reads your codebase, writes the changes, runs your test suite, reads the output, fixes what broke, and iterates until the job is done. Or until you tell it to stop, which is important to be able to do.

Originally released as Claude Dev (then rebranded as Cline), it’s become one of the most downloaded AI extensions on the VS Code marketplace, and the reason is that the agentic loop actually works. Not always perfectly, not without supervision, but well enough that it’s changed how a lot of developers approach tasks they’d previously either done entirely themselves or handed to a slower AI workflow.

What Cline Actually Does

Cline is an open-source VS Code extension that gives an AI model — your choice of Claude, GPT-4o, Gemini, or a local model via Ollama — the ability to take actions in your editor environment. It can read and write files, execute terminal commands, search the web, and even take screenshots of browser output. Each action is shown to you before it executes (by default), so you’re not watching your codebase get rewritten without oversight.

The task flow looks roughly like this: you describe what you want in the chat panel. Cline reads the files it thinks are relevant, proposes a plan, and asks for approval before making changes. It makes the edits, runs whatever tests or commands seem appropriate, reads the output, and continues until the task is complete or it hits something it can’t figure out. At that point, it asks you a clarifying question rather than guessing wrong repeatedly.

That approval model is one of the things that sets it apart. You can configure Cline to auto-approve file writes and command execution if you trust it — which speeds things up considerably — or keep manual approval for every step if you’re working on something sensitive. Most people find a middle ground: auto-approve reads and writes, manual approval for anything that touches the network or runs a deployment command.

The Model Matters a Lot

Cline’s quality as a coding assistant is essentially capped by the model you use with it. Claude Sonnet is the most common choice and the one Cline’s prompting is most clearly tuned for — it tends to plan coherently, produce clean diffs, and handle multi-file changes well. GPT-4o is a solid alternative. If you’re running a local model through Ollama, the smaller models (7B–13B parameters) work for simple tasks but struggle with anything requiring real architectural thinking.

You supply your own API keys, which means Cline’s cost is just your model usage. There’s no Cline subscription. For moderate use, running Cline on Claude Sonnet costs somewhere in the range of a few pence to a few pounds per session depending on the task complexity — substantially cheaper than a Cursor Pro subscription if you’re not using the IDE’s other features.

Where It’s Genuinely Better Than Copilot

GitHub Copilot and Cline solve different problems. Copilot is a completion and inline suggestion tool; it’s excellent at filling in boilerplate, autocompleting function implementations, and generating code at the line level. Cline is for tasks, not completions. “Write a migration script to move all users from table A to table B, run it against the test database, and verify the row counts match” is a task, and Cline handles it better than any inline completion tool could.

The web browsing capability is underrated. If Cline hits a problem it can’t solve from your codebase context — an obscure library bug, an undocumented API behaviour — it can search for the answer, read the relevant documentation or Stack Overflow thread, and incorporate that into its approach. That’s not something Copilot or a standard chat interface does in-context.

It’s also genuinely useful for refactoring across multiple files. Renaming a function and updating all its call sites, changing an interface and fixing all the implementations, migrating from one library to another — these are tasks that involve reading and editing many files in a coordinated way. Cline handles this more reliably than tools that operate one file at a time.

The Honest Limitations

Cline can go off the rails on complex tasks, particularly if it misunderstands the requirements early and confidently pursues the wrong approach. Keeping tasks focused and specific reduces this substantially — “add input validation to the createUser function” goes better than “improve the user management system.” For anything with meaningful complexity, checking in at natural points rather than letting it run to completion unsupervised is sensible practice.

The extension can be slow when working on large codebases, partly because it’s sending significant context with each model call. If your project is over a few hundred files, you’ll want to be explicit about which files are in scope rather than letting Cline decide.

Roo Code, a fork of Cline with some additional features around custom modes and more granular context control, is also worth knowing about — it’s maintained by a separate team and has built a following among users who want more configurability than the core Cline extension provides.

Getting Started

Search for “Cline” in the VS Code extensions panel, install it, add your API key in the settings, and start with a small task on a project where you’re comfortable with the model seeing your code. The GitHub repository at github.com/cline/cline has good documentation and an active community on Discord where you can find example prompts and configuration tips from people running it in production. It’s one of those tools that takes an hour to get comfortable with and then becomes part of your daily workflow.