【スマホ完結】通勤中にアプリが3つできた話（Claude Code on the web）A developer shares how they built three apps entirely on a smartphone during…

Anthropic's Claude Code began life as a terminal-based coding assistant, but it has increasingly moved toward the browser, and a recent first-person account illustrates what that shift makes possible. The developer describes building three small applications entirely on a smartphone while commuting, without ever opening a laptop. The story matters because it suggests that a meaningful slice of software development can now happen in short, mobile-friendly bursts rather than requiring a dedicated workstation.

The core enabler is Claude Code on the web, a hosted version of Anthropic's agentic coding tool that runs in a browser instead of a local command line. Rather than installing a runtime and cloning repositories onto a personal machine, the user connects a code repository, describes the task in natural language, and lets the agent plan, write, and run code inside a managed cloud environment. Because everything executes remotely, the client device mainly needs a browser and a network connection, which is why a phone becomes a viable control surface. The author frames this as the key unlock: the heavy lifting sits in the cloud, and the handset is reduced to a place for reviewing output and steering the next step.

In practice, a session like the one described appears to follow a loop familiar from other agentic tools. The developer issues a prompt, the agent proposes changes, executes them in an isolated sandbox, and reports back with results and diffs. The human then approves, corrects, or refines the direction. On a phone, this workflow leans heavily on reading and light typing rather than sustained coding, which fits the constraint of a small screen and an on-screen keyboard. The three apps produced this way are likely to be compact in scope, such as utilities or prototypes, since a commute imposes a natural time limit and mobile input remains slower than a full keyboard. The account is a single anecdote, so results will vary with project complexity and the quality of the prompts.

It helps to place this in the broader arc of Claude Code. The tool launched earlier as a CLI product aimed at developers who wanted an agent embedded in their existing terminal and editor setup, with access to files, shell commands, and version control. The web version extends that model into asynchronous, cloud-hosted sessions, and Anthropic has paired it with mobile access so tasks can be kicked off and monitored away from a desk. The underlying capability comes from Anthropic's Claude models, and the agent's usefulness depends on both the model's reasoning and the surrounding scaffolding that lets it read errors, run tests, and iterate.

This is part of a wider industry move toward what are often called asynchronous or background coding agents. Comparable offerings include OpenAI's cloud-based Codex agent, Google's Jules, GitHub's Copilot coding agent, and startup efforts such as Cognition's Devin, alongside editor-centric tools like Cursor. The common thread is delegating multi-step tasks to an agent that works in a sandbox and returns a reviewable result, sometimes several in parallel. Browser and mobile front ends are a logical next step for that pattern, because they decouple the work from any one machine.

Several caveats are worth keeping in mind. Granting an agent access to a repository and an execution environment carries security and code-review responsibilities that do not disappear on mobile; generated code still needs scrutiny before it is merged or deployed. Longer or more architecturally demanding projects are harder to manage from a phone, where reading large diffs and debugging subtle failures is cumbersome. Costs, rate limits, and the reliability of cloud sandboxes also shape how far this approach scales. The experience described is best read as a demonstration of feasibility rather than proof that phones will replace development machines.

Even with those qualifications, the account points to a genuine change in where and how coding can occur. If routine tasks, small tools, and early prototypes can be started and finished in fragmented time such as a commute, the practical definition of a development environment broadens considerably. Whether this becomes a mainstream habit will likely depend on how comfortable teams grow with reviewing agent output on the go and how much responsibility they are willing to hand to hosted agents.