Claude Code gets a new cloud-based planning mode
Anthropic has added a preview feature called Ultraplan to Claude Code, introducing a new workflow in which task planning happens in the cloud rather than tying up the developer’s terminal session. The change appears aimed at one of the most practical frictions in AI-assisted coding: planning can be useful, but it can also interrupt active development if it occupies the same working environment where the rest of the job needs to continue.
With Ultraplan, developers begin a planning job from the terminal, then move into the Claude Code web interface while the plan is built in the background. That means the terminal remains available for other work instead of being locked into a planning interaction. The feature is currently being offered as a preview for users who have enabled Claude Code on the web.
What changes in the workflow
According to the documentation summarized in the source material, Ultraplan differs from local planning in three main ways. First, planning runs in the background in the cloud. Second, users can comment on individual sections of a plan instead of replying to the whole proposal at once. Third, once the plan is complete, it can be executed either in the browser or back in the terminal.
Those differences matter because they turn planning into a more review-oriented process. In a local terminal flow, the plan is often treated as a single response that the user must accept or revise in one pass. A browser interface creates room for selective refinement. A developer can focus feedback on the parts that need correction without disturbing the rest of the structure.
The web interface also supports inline comments, emoji reactions, and revision requests. Those details may sound minor, but they suggest Anthropic is pushing planning toward a more collaborative document workflow rather than a transient chat exchange. For teams or individual developers working through complex tasks, section-level discussion could make AI-generated plans easier to edit, approve, and reuse.
Why planning is being separated from execution
The deeper significance of Ultraplan is architectural. AI coding tools are gradually splitting into distinct modes: planning, implementation, review, and execution. Anthropic’s latest feature treats planning as a separate cloud service that can proceed asynchronously, rather than as a blocking step inside the same local session where coding happens.
That reflects how many development workflows already function in practice. Planning is often iterative, collaborative, and document-heavy. Execution is immediate, stateful, and tied to local context such as a repository, shell, environment variables, and test runs. By moving planning into the browser, Anthropic is effectively acknowledging that those two modes benefit from different interfaces.
For developers, the practical gain is less waiting in the terminal. A plan can be generated and refined while coding, reviewing files, or running commands continues elsewhere. That could be particularly useful in larger repositories, multi-step tasks, or situations where the developer wants to inspect and challenge the plan before letting the model act on it.
Requirements and limits
Ultraplan is not available in every Claude Code setup. The reported requirements are a Claude Code web account, a GitHub repository, and Claude Code version 2.1.91 or later. The source also states that the feature does not work with Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry.
Those constraints indicate that Ultraplan currently depends on Anthropic’s own web-connected workflow rather than every enterprise or third-party deployment path. That may limit adoption among organizations that rely on managed cloud integrations or stricter internal infrastructure. It also suggests the feature is, at least for now, aimed first at users willing to operate inside Anthropic’s direct product surface.
Anthropic employee Thariq reportedly said on X that Ultraplan uses roughly the same number of tokens as the previous planning mode. If that holds broadly, the value proposition is less about raw model efficiency than about better ergonomics. The question is not whether planning becomes cheaper, but whether it becomes easier to integrate into how developers already work.
A sign of where coding assistants are headed
Ultraplan fits a broader pattern in AI coding tools: vendors are trying to reduce contention between the model’s workflow and the developer’s workflow. Early tools often treated the assistant as the center of the session. More recent designs are trying to make the assistant one process among many, able to work in parallel while the human keeps control of the environment.
That shift could prove important. Developers do not just want code generation. They want systems that can help decompose work, structure an approach, and then stay out of the way when needed. Moving planning into the cloud addresses that demand directly by giving the planning process its own place and timeline.
The browser-first design also opens the door to richer review behavior. Section comments and revision requests imply that AI-generated plans may increasingly become artifacts that can be discussed, versioned, and approved before execution. Over time, that could make planning outputs feel less disposable and more like formal development documents.
Preview status means the real test is still ahead
Because Ultraplan is still in preview, the key question is how well it performs in real workflows beyond early adopters. The concept is straightforward and useful: let the AI think in the background while the developer keeps moving. But the quality of the plan, the ease of revising it, and the friction of handing work back into the terminal will determine whether the feature becomes central or merely optional.
Even so, Anthropic’s move is notable because it treats planning as a first-class stage of software work. In that respect, Ultraplan is not just a convenience feature. It is a signal that AI coding platforms are becoming more modular, more asynchronous, and more opinionated about how development should be organized.
If that direction holds, the future coding assistant may look less like a chatbot in a terminal and more like a distributed collaborator with separate spaces for planning, discussion, and execution.
This article is based on reporting by The Decoder. Read the original article.
