Cline Is the Open-Source AI Coding Agent That Replaced Cursor in My Workflow, and You Are Paying $20 a Month for a Worse Version of the Same Wrapper

Hey guys, Mr. Technology here.

The most-installed AI extension in the VS Code marketplace in 2026 is not Cursor. It is not GitHub Copilot. It is an open-source project you have probably never heard of, and it is the sharpest rebuke to the "Cursor is the future of coding" narrative of the last eighteen months. It is not a new model. It is the wrapper, and the wrapper is open source, and the wrapper is the thing you are paying Cursor $20 a month for.

The project is Cline, Apache 2.0 licensed, ~50,000 GitHub stars, ~6,000 forks, 3,000+ contributors, v3.x as of June 2026, 5 million+ installs on the VS Code Marketplace. The same models. The same plan-then-act loop. The same browser automation. You do not pay Cline a single dollar — you pay the model provider for the API calls, exactly like on Cursor, except the wrapper markup is gone.

What Cline Actually Is

Cline is an AI coding agent that lives in your editor. You give it a task — "refactor this file to use async", "build a Next.js dashboard for /v1/users", "add error handling to every call site of this function" — and it executes end-to-end. It reads files, writes files, runs terminal commands, installs dependencies, calls the browser, iterates when tests fail, and reports back with a diff. It is the agent, not the chat.

The architecture is small. A TypeScript VS Code extension that wraps a model loop, a workspace-scoped terminal, a headless Chromium browser controller, and a diff viewer. The model loop is plain Anthropic / OpenAI / OpenRouter / Ollama / any-OpenAI-compatible-endpoint — the same SDK calls you would write yourself. The diff viewer shows pending edits before they hit the file system, with accept/reject per hunk. No proprietary backend. The model is the model. The wrapper is the wrapper.

typescript // The core loop, simplified const tasks = await plan(goal, context); for (const task of tasks) { const patch = await model.complete({ system: "You are Cline, a coding agent. Plan, then act.", tools: [readFile, writeFile, runCommand, browse], messages: history, }); if (patch.type === "edit") { await applyDiff(patch.path, patch.diff); history.push({ role: "assistant", content: Edited ${patch.path} }); } }

Plan, execute, observe, iterate. The wrapper does not need to be smarter than the model. It just needs to be honest about what it is.

Plan/Act And The Browser

Cline 2.x introduced the Plan/Act split that Cursor has been trying to clone for twelve months and still gets wrong. In Plan mode the agent reads the codebase, drafts a multi-step plan, and waits for approval. In Act mode it executes with full tool access — file edits, terminal, browser, MCP servers. The split is the right abstraction: planning and execution are different cognitive loads, and forcing them into a single turn is what makes every other agent hallucinate file paths and skip the test step.

The browser integration is the second piece Cursor does not have. Cline can browser_action to navigate, screenshot, click, type, scroll, extract, and answer questions about any page. The use cases that pay rent: "scrape this dashboard into a CSV", "find the rate limit in API X's docs", "log in to staging and reproduce the bug from the ticket". Cursor is not in the browser. Cline is.

What Is Actually Wrong

Three things. Plan mode is not free — on long multi-file refactors the plan can run 8 to 15K tokens of pure scaffolding before the act loop even starts. The browser tool is slow on the same Anthropic-hosted models Cursor uses — a single browser_action round-trip is 4 to 8 seconds on Claude Opus 4.8, and a page with twenty questions is a 90-second loop. The cloud offering is the obvious trap — the path from "I ran it locally" to "I want this in CI for 200 engineers" goes through the paid product, the same pattern as Pipecat, Aider, every open-source agent that ships a hosted tier. The extension is genuinely open. Read the FAQ.

The Take

Cursor charges $20 per month for Pro. Copilot charges $10 to $39 per user per month. Cline charges you nothing. The same model. The same Anthropic API key. The same plan/act loop, with the browser tool that Cursor does not have, and the open-source extension that 5 million engineers have installed. You are paying for the wrapper. The wrapper is the product. The wrapper is open source. The wrapper is at 50K stars and 5 million installs.

If you are using Cursor, you are paying $240 a year to use an LLM you can call directly. Stop. Install Cline. Plug in your Anthropic key. The diff viewer and the browser tool do the things you actually needed Cursor to do. The math is the math. The model is the model. The wrapper is free.

— Mr. Technology

Repo: github.com/cline/cline — Apache 2.0, ~50K stars, ~6K forks, 3,000+ contributors, v3.x as of June 2026. 5M+ installs on the VS Code Marketplace. TypeScript VS Code extension. Model-agnostic: Anthropic, OpenAI, OpenRouter, Groq, Ollama, any OpenAI-compatible endpoint, plus local models via LM Studio and llama.cpp. Plan/Act mode shipped in v2.x. Browser automation via headless Chromium, integrated as a first-class tool alongside file edits, terminal, and MCP. MCP client and server in v3.x. Cloud offering: cline.bot (managed runtime for teams). License: Apache 2.0.