JetBrains has released a new version of Air, its parallel-workspace tool for running AI coding agents, adding support for GitHub Copilot, OpenCode, Pi, Cline, and other agents built on the Agent Client Protocol, alongside IntelliJ-powered navigation for Java and Kotlin and Docker task support on Windows.
Air's core idea is giving coding agents separate workspaces to run tasks in parallel, with a review step before any of an agent's changes reach the actual codebase. Until this release, that workflow was limited to the agents JetBrains bundled directly into the product, including Claude, Codex, Gemini CLI, and Junie. The new release opens that workflow to any agent that speaks ACP, an open protocol developed jointly by JetBrains and Zed that gives coding agents and development environments a shared way to communicate, so Air can connect a supported agent to its workspace without needing a bespoke integration built for that specific tool.
That matters most for organizations that have standardized on a single approved AI coding tool and don't permit alternatives. Previously, a company that had only approved GitHub Copilot effectively locked its engineers out of Air, since Air had no way to connect to the Copilot access the company had already provisioned. Air now connects to the GitHub Copilot CLI through its ACP server mode, letting engineers use their company-managed Copilot access while still getting Air's parallel workspaces and review workflow layered on top.
The same principle extends to other ACP-compatible agents. A model is only one piece of a coding agent; the harness around it, the software that gathers context, calls tools, manages the task, and turns model output into code changes, differs meaningfully between tools and supports different providers. Air users can now pick the harness that fits their situation, running OpenCode with a preferred cloud provider or Pi with a model running entirely on their own machine, with the available models and agent-specific features depending on what that particular harness exposes through ACP.
Local-model support was among the most requested additions from users. A model running locally through a runner like Ollama or LM Studio can now connect through an ACP-compatible agent, which Air then connects to in turn, letting developers work entirely offline and choose whichever model best fits their codebase or task.
On the language side, Air now supports Java and Kotlin language intelligence, including mixed Java and Kotlin projects, powered by the IntelliJ IDEA code engine in beta. While creating a task, developers can jump to definitions, find usages, search symbols, and follow code paths directly inside Air, giving them more precise context to hand an agent before it starts working. Once an agent finishes, Air highlights errors and warnings in the affected code, letting developers catch problems before accepting changes rather than waiting for a failed compile.
Air can now run agent tasks inside Docker containers on Windows, matching support that already existed on macOS, for developers who want dependencies and agent commands isolated from their host machine rather than running directly on it; Docker Desktop is required. Smaller quality-of-life additions round out the release: clicking a file name when an agent proposes an edit now opens the complete change in a dedicated tab rather than a short chat snippet, Air displays context-window usage and token counts for Claude Agent so developers can tell when a long task is approaching its limit, Claude Fable, Opus, and Sonnet 5 gain an "xhigh" reasoning effort setting for tasks that need it, and a new macOS setting prevents the machine from sleeping mid-task.
The update is available now for download at JetBrains' site or through JetBrains Toolbox.