Junie The Anti-Vibe Coding IDE

AI coding tools often fail hardest when a developer already has a real repository to maintain—because they tend to “go wild,” generate large diffs, and avoid the kind of planning, review, and control professionals expect. Junie, from JetBrains, is positioned as an “anti-vibe coding” Smart Coding Agent built into JetBrains IDEs so it can work inside the language- and framework-specific workflows developers already use.

Rather than acting like a standalone plugin that blindly writes code, Junie is described as fundamentally baked into JetBrains IDEs such as IntelliJ IDEA, PyCharm, PhpStorm, WebStorm, and RubyMine. That integration matters: Junie can scaffold and operate in a way that matches the project’s architecture—compared to how an Android app’s structure maps to a specific IDE setup. The result is closer to pair programming than prompt-to-code automation. Developers can converse with the agent while it understands the codebase context, then execute changes through an iterative loop: plan what to do, act by writing or optimizing code and tests, verify, and refine.

A key strength highlighted is “ask mode,” designed for working in existing codebases. Instead of immediately generating code, Junie can scan the repository and answer targeted questions about frameworks, setup details, and how components interact. From those answers, it can produce project-specific guidelines saved to a markdown file—an “agent’s rules” document that helps keep subsequent work aligned with team standards. This is paired with an execution mode that can take tasks, run terminal commands, and make changes while still keeping the developer in the loop.

Control and auditability are treated as first-class features. Even in “brave mode,” where Junie attempts to fix issues by writing code, it can pause for confirmation before running certain terminal commands. The workflow also emphasizes transparency: each run includes a plan, a clear breakdown of what changed (added vs. deleted code), and a diff view that makes rollback straightforward. Developers can then commit changes to Git, push to branches, and generate change logs based on recent commits.

The transcript also spotlights practical configuration: Junie supports guidelines, MCP integration (by adding an MCP server with an API key), and model switching between options such as GPT-5 and Sonnet 4.5. Tasks can be executed with specific context (including “context seven”) so the agent can use the tools exposed through MCP.

A live example centers on converting a demo using Gemini’s Live API/BiDi into a NextJS version, adding features for a live customer support console (including telephone-style query handling), and then fixing UI issues caused by the conversion—described as a CSS linking problem. The agent’s output is framed as usable software rather than eye-candy code generation: the goal is the least-change approach that completes the job.

Overall, Junie’s pitch is that success shouldn’t be measured by raw lines of code produced. It’s measured by controlled, reviewable modifications that help developers onboard faster, learn new frameworks and languages, and ship working software without burning excessive tokens—while still behaving like a collaborative coding partner inside JetBrains IDEs.