Claude Code for web is the hosted, asynchronous form of Claude Code, launched by Anthropic in October 2025 and available through the Claude web interface and as a tab in the Claude iPhone app. What the source establishes outright is that it runs in a container managed by Anthropic; its identification of what runs inside that container is offered as the author’s own inference from the outside — as far as he can tell, the Claude Code CLI wrapped in a container and configured to skip permission prompts, which he reports appears to behave exactly like the local tool.
It belongs to the category of asynchronous coding agents, and the source positions it as Anthropic’s counterpart to OpenAI’s Codex Cloud and Google Jules, with a very similar shape: point it at a repository, give it a prompt, and collect the result as a branch rather than watching it work.
Capabilities
A run is set up by pointing the agent at a GitHub repository, choosing an environment, and supplying a prompt. While a run is in progress, further prompts can be sent and are queued for execution after the current step completes. When the run finishes, the agent opens a branch on the repository carrying its work and can optionally open a pull request. A “teleport” feature copies both the chat transcript and the edited files down to the local Claude Code CLI, so a session started in the hosted environment can be taken over locally.
The environment choice is the security-relevant control, and the source describes a range from fully
locked down, through a restricted allow-list of domains, to a configuration allowing domains of the
user’s choosing including * for everything. The only run whose environment the source specifies is
at the open end of that range — a benchmark against a private repository under a custom *
allow-list, chosen because that project involved no secrets or source code needing protection. Separately it states that running
with no network access leaves nothing to worry about, and records unease about the default
allow-list of the middle setting.
Notes
The source’s assessment is that the pull requests it produces are indistinguishable from the local CLI’s — it reports that the same prompt would likely have given the same result on a laptop — and locates the product’s value entirely in convenience: a hosted container managed by Anthropic with a web and mobile interface over it. The examples it gives are a small single-file tool, a README correction, and a multi-scenario Python templating benchmark complete with charts, the last of these entered on a phone keyboard.
On cost, the source declines to generalise, noting it was using a plan provided by Anthropic for testing and estimating only its own daily CLI spend via an unofficial tool.
Related
Claude Code, Sandboxing, Lethal Trifecta, Google Jules, OpenAI Codex, Claude Code for web—a new asynchronous coding agent from Anthropic