JetBrains Air Is a Control Layer, Not a New Model

JetBrains did not ship a model on Sept. 22. It shipped a name for the mess around the model. Paul Sawers at The New Stack called JetBrains Air one of the most significant steps in the company’s 26-year history, and he put the bet in the headline: the IDE still matters. I buy the second half more than the first. Significant is a press word. The IDE question is the actual product question.

Unite.AI’s writeup of the same day is the cleanest inventory. Air is an open system spanning the IDE, team delivery, and organizational governance. Junie, JetBrains’ own coding agent, is supposed to sit on every surface. That is a packaging announcement. It is also the first time the company has said, in one place, that the agent is not the product you are buying. The product is the layer that decides which agent runs, where the files live, and who can see the bill.

We already looked at Coder’s Agent Relay, which keeps Claude Code’s loop at Anthropic and moves the filesystem onto machines you own. We looked at Goose, which lets you swap models and does not swap out the token invoice. Claude Code on Ollama is the zero-invoice path, with the limits that come from a local model. Air is a fourth shape. It does not claim to move inference. It claims to be the desk where several agents, including ones JetBrains does not make, get pointed at work and then reviewed.

What JetBrains Air lets you open

Sawers’s recap is the part I trust for availability, because it separates the March preview from the September rename. JetBrains put Air in public preview in March as a standalone agentic development environment, at first on macOS. In that app you could run Claude Agent, Codex, Gemini CLI, and Junie side by side. At the same time Junie got a CLI, so the agent was no longer stuck inside the IDE. Terminals, CI, other editors.

The September move does not delete that app. Sawers says the desktop IDE is still there, on macOS, Windows, and Linux, plus a browser version for organizations. “Air” now does two jobs. It is still the name of that agent-first desktop app. It is also the banner over a set of products that used to have other names.

Where you meet it, per The New Stack: inside JetBrains IDEs, in the browser, and from the command line through something called Air Gateway. Gateway is the piece that pulls terminal agents such as Claude Code and Codex into Air. If you already live in a terminal and you do not want a new editor, Gateway is the sentence to test. If the test is “does my existing Claude Code session show up as a task I can review,” I have not run that test. The announcement says the pipe exists.

Sander Kleijn at Techzine, also Sept. 22, describes the same week as a consolidation. Six months of public experiments, including JetBrains Central, get one umbrella. Central continues, under the name Air Governance. Junie, which plans and runs multi-step tasks in the IDE, is supported on all Air interfaces. Consolidation is not a feature. It is a map. The map matters if you were trying to buy “the JetBrains agent thing” in April and found three product names.

Three products, and a calendar that does not match the blog posts

Unite lists three products in the Sept. 22 introduction.

Air in JetBrains IDEs is the experience for pointing agents at work and checking what they did, without leaving IntelliJ or whatever cousin you already pay for. This is the “the IDE still matters” bet in concrete form. The agent can live in a terminal. The review, JetBrains is arguing, still wants the editor that knows the project.

Air Teams coordinates delivery workflows that include both people and autonomous agents. Sawers’s captions show shared projects and automations in a browser. That is a team board, not a model.

Air Governance used to be JetBrains Central. Unite’s job description is policy, visibility, auditability, cost management, and accountability for AI-assisted work. Sawers is more specific and more cautious: which models and agents a developer may use, permissions, spending limits, usage across teams. He also says those governance controls are only in an early access program for now.

Hold those two sentences together. The launch post names Governance as a product. The New Stack says the controls are early access. IT Europa’s Sept. 23 note from an Amsterdam briefing makes the calendar worse, in a useful way. CEO Kirill Skrygan’s roadmap, as that piece reports it, puts Air Teams, Air Governance, and Air Credits out next month, Air Mobile before year end, and a subscription price for the full Air bundle next January, after the pieces launch individually this year.

“Announced Sept. 22” and “next month” do not have to conflict. Both can be true if Sept. 22 was a naming event and the SKUs trail it. What you should not do is file a purchase order for “JetBrains Air” as if it were one GA product with a price. IT Europa says current access is a JetBrains AI subscription, AI Pro or AI Ultimate, or your own API keys. That is the access path for the agent surfaces that exist. It is not the January bundle price, which does not exist yet.

Unite adds one availability fact that is easy to miss. Cloud runs are already on for some customers in JetBrains IDEs and the browser. Everyone else is a gradual rollout over the coming months. “Some customers” is not a region list. If your admin console does not show it, you are in the later group until someone tells you otherwise.

ACP is the part that is not a lock-in slide

The Agent Client Protocol is the piece I would actually read, because it is the only part of this launch that other editors can use.

Unite says ACP is an open protocol for how an IDE and a coding agent talk, developed in the open by JetBrains and Zed, with a stated goal of no vendor lock-in. The harness it is supposed to expose includes planning, logic, tools, model routing, and observability, not a bare chat socket. An ACP Registry lets you find compatible agents and run them inside JetBrains IDEs.

The agent list on that page is long enough to be a claim you can check: Junie, Gemini CLI, GitHub Copilot, Codex, Cursor, Mistral Vibe, OpenCode, Kimi CLI, Qwen Code, Factory Droid, Cline, Kiro CLI. Clients named: JetBrains IDEs, Zed, and Neovim via a plugin. The product page, again per Unite, says Air works with Claude Agent, Codex, Junie, Copilot, OpenCode, and any agent on ACP. Mobile is marked coming soon.

IT Europa’s provider list is looser. The reporter writes that Air is designed to work across services from Anthropic, Amazon, Google, Microsoft, OpenAI, or SpaceX, as examples. I would not put SpaceX on a procurement spreadsheet because a trade-press sentence used it as an instance. Use Unite’s named agents when you are deciding whether your current tool is on the bus. Use IT Europa for the CEO’s product philosophy, not for the compatibility matrix.

Techzine has the company line that makes ACP more than a logo wall: “Supporting multiple models and agents is the floor, not the ceiling.” JetBrains does not expect the market to collapse to one agent soon, and it is treating multi-vendor support as a design rule. That is the opposite of a model launch. A model launch wants you on one endpoint. A control layer wants every endpoint to show up in the same review queue, with a spend cap.

If ACP is real in the way Zed and Neovim being clients implies, Air is not the only door. You could care about the protocol and ignore the JetBrains bundle. I have not verified a Neovim session against the registry this week. The reviewable claim is that JetBrains published a client list that includes editors it does not own. That is either the start of a real bus or a list that will rot. Check the registry against one agent you already run before you believe the wall of names.

Isolated tasks, and the code that looks finished

IT Europa’s feature list is the closest thing to a review checklist, and it is still a briefing, not a lab note.

Multi-agent orchestration: Claude Agent, Codex, Gemini CLI, and Junie side by side, without stepping on each other. Parallel task execution: hand different jobs to different agents at once. Isolated environments: local workspaces, git worktrees, or Docker. A code review interface that uses project intelligence, so you look at the diff before it merges. macOS, Linux, Windows.

The git worktree item matches a caption in Sawers’s piece, a new worktree task in the desktop app. Isolation is the grown-up version of “the agent edited the wrong branch.” If the task dies with the worktree, you can throw the attempt away. If the task can see your main checkout and your cloud credentials, you have recreated the problem Coder’s product is trying to box in, only inside an IDE vendor’s desktop app. There is a feature name. Treat “isolated” as a claim to test with a junk repo, not as a security boundary you can show an auditor yet.

Skrygan’s quote is the one that earns the launch, and it is not about JetBrains. “AI can produce code, but organisations still have to produce software. Agentic development is changing how software gets made, but it hasn’t changed what it costs to be wrong.” Then the distinction that matters more than the product grid. Obviously wrong code gets caught. The expensive code is almost right: plausible, able to pass a shallow check, carrying a bad assumption that shows up when the fix is no longer cheap.

That is the review you are actually buying, if you are buying anything. Air’s code review UI is the company’s answer to its own CEO. Whether the UI catches an architectural inconsistency, or only a lint-shaped diff, is the test I cannot run from a press briefing. Techzine reports JetBrains’ own diagnosis in plainer ops language. Developers are picking up agents faster than organizations can build the support around them. Context, coordination, governance, and cost control are behind. The costs show up later, in reviews, rework, security, infrastructure, and spend.

Air Governance is aimed at that list. Early access, per Sawers. Next month, per the Amsterdam roadmap. If your reason for looking at Air is the spend cap and the audit log, wait until you can click those screens, or get into the early access program and write down what is actually enforced. A renamed Central console that cannot yet deny a model is a slide.

What I would do with it this week

If you already pay for a JetBrains IDE and an AI subscription, open whatever Air surface the September build put in the IDE and run one agent you already know, through ACP if that agent is on the list. The question is not whether Junie is smart. The question is whether a Codex or Claude Code task shows up as a reviewable unit, in a worktree, with a diff your existing inspections can read. If it does, the control layer is real enough to keep. If it is a chat panel with a new title, you already own that.

If you wanted inference to stay in the building, this is the wrong ticket. Nothing in these four writeups moves the model weights onto your GPU. Ollama is still that article. If you wanted the agent to stop writing on the laptop that has production kubeconfigs, read the Coder piece again. Air’s Docker and worktree isolation might get you part of the way on a developer machine. It is not a described VPC boundary.

If you are an admin being asked to “standardize on Air,” ask which of the three products is in the contract, whether Governance can deny an agent today or only next month, and what the January bundle price is supposed to replace. IT Europa says the company has a limited consulting practice and about ten regional resellers, and that Skrygan expects to need more integrators after this launch. That is a tell. The software is ahead of the people who would install it for a bank. You will be installing it yourself, or waiting.

I am not scoring Junie against Claude Code. This week the company did not ask for that score. It asked you to treat the agent as a replaceable worker and the IDE, plus a governance console that is not fully out, as the system of record. The IDE half you can open. The console half you should not pretend is finished because the blog post used the present tense.