Cline, a free VS Code coding agent, gets a 7.5/10 review
A dev.to review of Cline, an open-source assistant that edits many files at once, shows what it does well, what it costs, and how to supervise it.
The short version
- Cline (formerly Claude Dev) is an open-source agent that runs inside VS Code and can edit files, run commands and use a browser.
- It shows its plan and a diff for each change, and you approve before anything is written.
- The extension is free, but you pay your chosen model provider; the review puts a typical Claude Sonnet session at $0.50-2.00.
- Reviewer's tests: auth for a Next.js app in 12 minutes, 85% of generated React tests passing first run.
- Listed downsides: less polish than Cursor, confusion on very large codebases, and API costs on long runs.
Why it matters
Coding agents that edit many files at once can save real time, but they can also spread mistakes widely. Cline's approach of showing every step and asking for approval is a useful template for supervising any AI tool. Because it is free and works with many models, it is also a low-risk way to try an agent without switching editors.
What Cline actually is
Cline, which used to be called Claude Dev, is a free extension for VS Code, a widely used code editor. As dev.to reported in a review originally published at pikvue.com, it is an agent: software that takes a goal, breaks it into steps and carries them out, rather than just suggesting the next line as you type.
You describe a task in plain language. Cline reads your project, drafts a plan, then creates and edits files, runs terminal commands, reads error messages and tries again. It can also open a browser to check documentation.
The review calls it the open-source answer to the agent mode in Cursor, a rival editor with AI built in. This is one reviewer's opinion, not a benchmark.
Supervision is the selling point
The review's central praise is visibility. Each prompt, tool call and file change is part of the session record, so when something goes wrong you can find the specific step instead of guessing what the model did.
Cline also shows a diff (a side-by-side view of what will be added or removed) before it changes anything. You approve each write, or you can loosen the settings and let it run on its own. The reviewer treats the approval step as the line between an agent you supervise and one you merely hope is right.
Because the extension's code is public, security-minded teams and maintainers can inspect or copy it. That matters if you do not want to depend on a tool you cannot examine.
Bring your own model, and your own bill
Cline does not include an AI model. You connect your own, through an API, which is the doorway that lets one program send requests to another service. The review lists Claude, GPT-4, Gemini and DeepSeek, plus local models that you run on your own hardware.
So the extension costs nothing, but the model usage does. The review puts a typical session with Claude Sonnet at $0.50-2.00, about the same with GPT-4o, and effectively free with a local model like DeepSeek Coder once you have the hardware. It also warns that long, autonomous sessions can make API charges pile up.
The upside is control over cost and privacy. A team can route a smaller or self-hosted model through the same tool without changing its workflow.
What the reviewer's tests showed
The review describes three tasks. These are the author's own reports, not independent measurements.
- Adding login to a Next.js app: 12 minutes, with one correction needed for a wrong redirect address. The reviewer says it would take 30+ minutes by hand.
- Writing tests for a React component: 8 minutes, with an 85% pass rate on the first run. Cline fixed the 2 failing tests on its next pass.
- Refactoring a Python API: 15 minutes to split one large route file into separate modules, with imports updated correctly.
The listed weaknesses are a rougher finish than Cursor, confusion on very large codebases, and the cost of long sessions. The final score is 7.5/10.
How to use an agent without trusting it blindly
The review recommends starting with small, well-defined tasks. Read Cline's plan as you would a colleague's proposal: is the approach right, and is it touching files it should not? Only then let it draft the change.
After a batch of work, run your tests and a linter (a tool that flags code mistakes and style problems), then read the final diff once more before sharing it. The goal is to treat the agent's output as a proposal you have checked.
This slows things slightly, but it moves effort from cleaning up after bad changes to catching them early.
What to watch
- Whether Cline's reported results hold up on larger, messier codebases than the review's examples.
- How real-world API costs compare with the review's $0.50-2.00 estimate when sessions run long on auto-approve.
Try it
Before letting any coding agent loose, commit your current work so you have a clean restore point, then read each proposed diff before approving it.
Sources and links
- Reportingdev.to: the article we sourced
First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.