JetBrains Junie reached general availability on June 17, 2026, after months of iteration following its early 2025 prototype. This release puts a native agent into IntelliJ IDEA, PyCharm, WebStorm, Rider, and GoLand. Junie achieved a 61.6% resolution rate and a 72.7% pass@5 rate on the SWE-Rebench leaderboard, which places it ahead of GitHub Copilot’s last documented agent-mode score of 56%. While Cursor’s Agent Loop with GPT-5.2-codex scored 67.4% on a SWE-bench Verified subset, Junie provides a different workflow by integrating directly with the existing JetBrains semantic index and debugger.
Getting Junie running in your project
You can start using Junie in the IDE or via the Junie CLI in your terminal on macOS, Linux, or Windows. To use the CLI, you must first install it and navigate to your project root directory to run the junie command. After the first launch, you select an authentication method: a JetBrains account, a Junie API key for usage-based billing, or your own API key through a "Bring Your Own Key" (BYOK) setup. If you use BYOK, you connect to providers like OpenAI, Anthropic, Google, or xAI, and those providers bill you directly.
Setting up project-specific rules helps Junie stay aligned with your team conventions. You should create a .junie/guidelines.md file in your project root to define coding standards. For those who prefer high-level autonomy, the Brave mode setting allows Junie to execute terminal commands without asking for permission every time. You can toggle this via the /brave command or the Ctrl+B shortcut. Users can also enable Sandbox mode with junie --sandbox to restrict the agent to specific directories.
| Feature | Junie CLI | Claude Code |
|---|---|---|
| Supported Models | Anthropic, OpenAI, Google, Grok | Claude models only |
| Context Window | ~200k tokens | ~400k tokens |
| IDE Integration | Deep JetBrains integration | Terminal-first |
| BYOK Support | Yes | No |
Managing costs and model selection
JetBrains offers several ways to pay for Junie, and the credit system causes significant frustration for some users. The free tier includes 5 AI credits with no card required, while paid tiers scale from $8.33 per month for 10 credits to $25 per month for 35 credits. I found the recent shift toward strict credit consumption for premium models like ChatGPT 5 to be a major drawback for developers who expected more predictable costs. One user reported that their 35 credits dropped to 10 after only a few hours of normal use, leading them to cancel their subscription.
If you want to control costs, you can choose between different models in the Junie interface. You can select Claude Opus 5.5, Gemini 3.8 Flash, or GPT-6.1 SOL. For those wanting to avoid cloud costs entirely, Junie Local provides an on-device mode that requires an M5 Mac with 64 GB of memory. This local mode lets you run Qwen3.8-3.6-27B-blend, which generates 71% fewer output tokens than the standard Qwen3.8.
Using the agent for complex tasks
Junie handles multi-step tasks by using its "Plan mode," which you enter with Shift+Tab. In this mode, Junie creates a structured document in .junie/plans containing technical design and delivery stages for your review. You must approve the plan before Junie begins writing code. This prevents the agent from wasting tokens on a path you might reject.
The agent also performs debugging by interacting with the IDE’s actual debugger instead of just adding print statements. You can ask Junie to "Stop at line 42 in Main.kt" or "Why is x null?" to have it inspect stack frames and variable states. This feature requires IntelliJ IDEA 2026.1.1 or newer and the enabled MCP Server and Debugger MCP Toolset plugins. For frontend developers, Junie can scaffold components and run automated tests to verify the UI. Will the current credit consumption model remain sustainable for individual developers as models become more expensive? To start, install the plugin and run your first prompt.
