The landscape of AI-assisted development has taken a significant leap with the latest update to Air, the platform that enables running coding tasks in isolated, parallel workspaces. The new version not only expands compatibility with external agents such as GitHub Copilot, OpenCode, Pi, and Cline, but also introduces support for local models and improves navigation in Java and Kotlin projects powered by the IntelliJ engine. This move addresses two of the most recurring demands from development teams: the ability to use the AI tool already approved by the company while retaining the flexibility to choose the model or harness best suited for each task.
To understand the real impact of this update, it is worth analyzing how Air transforms the traditional workflow. Until now, the platform offered separate workspaces for agents like Claude, Codex, Gemini CLI, or Junie, allowing each to run tasks in parallel and letting the developer review proposed changes before merging them into the repository. With the new version, this same flow extends to any agent compatible with the Agent Client Protocol (ACP), an open protocol developed jointly by JetBrains and Zed. ACP establishes a common language between coding agents and development environments, eliminating the need for custom integrations for each tool.
The most immediate consequence is that companies that have standardized on GitHub Copilot as their only approved AI tool can now connect it to Air without changing providers. Simply activate the ACP server mode of the Copilot CLI so that Air recognizes it as another agent, inheriting all the access and model management that the company already had configured. This approach not only reduces friction with security and compliance departments but also allows developers to benefit from Air's orchestration — parallelism, change review, task isolation — without giving up their corporate tool.
Flexibility does not end there. Air also opens up to other agents like OpenCode and Pi, and allows developers to choose the harness they prefer. The harness is the software that wraps the model: it gathers context, calls tools, manages the task, and converts model responses into code changes. Each harness approaches this process differently and supports different model providers. With Air, you can use OpenCode with a cloud provider or Pi with a local model, and the platform handles communication via ACP.
Precisely, support for local models has been one of the most anticipated features. It is now possible to connect an ACP-compatible agent to a model running on your own computer using managers like Ollama or LM Studio. This allows offline development, choosing the model that best fits the project or environment, and keeping data on the local machine—especially valuable in contexts where privacy or security policies limit cloud services. At Q2BSTUDIO, as a company specialized in custom software development, we see this capability as an opportunity to integrate AI assistants in controlled environments without relying on external APIs, something more and more clients request in their digital transformation projects.
In the realm of language intelligence, Air now incorporates the IntelliJ IDEA code engine (in beta) to offer navigation and diagnostics in Java and Kotlin projects, including mixed projects. While creating a task, the developer can jump to definitions, find usages, locate symbols, and follow code paths directly in Air. This helps understand the code the agent will modify and provides more precise context. Once the agent finishes, Air highlights errors and warnings in the affected Java and Kotlin code, so the developer can inspect problems before accepting changes, without opening the IDE or waiting for compilation failure. For teams working with statically typed languages, this integration drastically reduces review time and prevents subtle errors from reaching production.
Another important improvement is the ability to run tasks in Docker containers on Windows, previously only available on macOS. This allows dependencies and agent commands to run in an isolated container, with Docker Desktop required. In corporate environments where Windows is the standard operating system, this functionality is key to ensuring reproducibility of development environments. Additionally, the update includes several enhancements to control during agent runs: you can inspect the complete proposed change before granting permission — clicking the file name to open the full edit in the 'Proposed Change' tab —, context window usage and token counts are shown for the Claude agent, and an 'xhigh' effort level is added for tasks requiring more reasoning with Claude Fable, Opus, and Sonnet 5. A setting to prevent the Mac from sleeping during long tasks has also been added, something developers working with batch processes will appreciate.
From a broader perspective, this update consolidates Air as a platform that not only competes with other AI agent solutions but also offers a unified framework for teams that must deal with multiple tools, security policies, and compliance requirements. At Q2BSTUDIO, where we develop cloud AWS/Azure projects and AI solutions for our clients, we particularly value the ability to integrate external agents without renegotiating contracts or abandoning already validated tools. The combination of local models, ACP compatibility, and the power of the IntelliJ engine positions Air as a solid option for teams looking to automate parts of the development cycle without sacrificing control or quality.
Of course, cybersecurity remains a critical factor. Being able to run agents in Docker containers and work with local models reduces the attack surface by minimizing connections to external services. Additionally, the change review flow before integration acts as an extra barrier against malicious code or unwanted modifications. In projects where we handle sensitive data, such as those requiring advanced cybersecurity, this architecture aligns perfectly with DevSecOps best practices.
Regarding business intelligence, although Air is not directly oriented toward data analysis, the ability to connect agents that can generate reports or perform code transformations can be integrated into broader BI / Power BI flows. For example, an agent that accesses a database and generates SQL queries could be orchestrated from Air to then feed a dashboard. The flexibility of ACP allows these use cases to be explored without developing specific connectors.
Finally, it is worth noting that Air is available for Mac, Linux, and Windows, and can be downloaded from air.dev/download or updated via JetBrains Toolbox. The company invites users to try the new version with existing agents and models and share their experience. For development teams already using AI tools or considering adopting them, this update represents a turning point: for the first time, it is possible to centralize the management of heterogeneous agents in a single workflow, maintaining developer autonomy and company governance.
In short, Air positions itself as an orchestration layer that solves one of the thorniest problems of AI adoption in software development: tool fragmentation. By opening up to external agents and local models, it not only democratizes access to artificial intelligence but also offers a pragmatic path for companies to advance their digital transformation without replacing their entire development ecosystem. At Q2BSTUDIO, where we help organizations implement automation solutions and custom software, we see Air as a natural ally for projects that require agility, control, and security.
