Cline - the most underrated AI code editor

Cline (spelled “client” throughout) is a VS Code extension for building real software with AI agents—starting from a simple prompt and ending with working, interactive apps—while staying fully open source. That combination matters because it gives beginners a guided “plan then act” workflow, and it gives power users transparency and control: the agent’s prompts are visible, the code can be forked, and improvements can be contributed back to the main repository.

The setup begins with installing VS Code, then adding the Cline extension. After installation, users can either start with free access or paste their own API key to unlock more capable models. Cline supports multiple API providers (including OpenRouter, Anthropic, and Amazon Bedrock), but the workflow presented relies on OpenRouter: create an account, add credits, generate an API key, and treat that key like a password.

A key differentiator is how Cline runs inside the VS Code ecosystem rather than as a standalone app. Cursor and Windsurf also sit on VS Code, but Cline’s open-source status is positioned as the main advantage: developers can inspect how it works, view internal prompt logic, fork it, and submit pull requests. Recent updates are highlighted as practical upgrades rather than marketing—improved Gemini caching and transparency to reduce token usage, more flexible “enhanced checkpoints” across task workflows, and a command (/new rule) that can generate project standards (described as producing something like a Cursor rules file) autonomously.

The workflow itself is built around two modes. In Plan mode, Cline first creates an implementation plan and asks questions; only after the plan is accepted does it switch to Act mode to execute changes. The transcript demonstrates this with a one-shot chess game: Cline produces an architecture overview (including a visual graph), then—after enabling “YOLO mode” (auto-approve)—it proceeds without repeated permission prompts. YOLO mode is framed as higher-risk autonomy, with controls for what actions are allowed (reading and editing files, executing save commands, optionally executing all commands, and enabling browser automation) and a cap on how many autonomous requests it can make before asking again.

The most striking capability shown is browser automation. After generating code, Cline runs a headless browser to verify the app works, then reports task completion. The chess game is tested in-browser with full rules behavior, including check and legal responses. The cost and context transparency are also emphasized: users can see token in/out, context window usage, and API cost (the chess example lands at about $2.5 using Claude 3.7).

To prove it’s not a one-off, Cline is used again to one-shot a terminal-based Minesweeper. In Plan mode it spends substantial tokens reasoning before writing code (23,000 tokens before any lines), then in Act mode it generates the full app and runs it in the terminal. The result is an interactive Minesweeper with configurable board size and mine count.

Beyond the demos, the transcript pushes a mindset shift: “vibe coding” isn’t just turning off thinking—it’s “vibe learning,” using AI to debug and understand errors. The practical takeaway is that non-programmers can build interactive games and full apps by writing a single prompt, accepting the plan, and letting Cline handle the implementation—while still having enough visibility to learn from what happens.