Claude Code for web–a new asynchronous coding agent from Anthropic (simonwillison.net)

Anthropic today launched Claude Code for web, a hosted, asynchronous coding agent that mirrors their Claude Code CLI experience in a browser and iPhone app. Users point it at a GitHub repo, choose an environment (fully locked down, allow-listed domains, or “*” for full access), send a kickoff prompt and queue follow-ups while the agent works step-by-step. It runs in a containerized runtime (basically the CLI wrapped with a web UI and a “teleport” feature that copies chat transcripts and edited files to your local Claude Code CLI), opens a branch with its changes and can create a PR. Practical demos included a complete benchmark comparing MiniJinja vs Jinja2 across Python 3.14/3.14t scenarios, with scripts, results and matplotlib charts produced end-to-end by the agent. The notable technical move is Anthropic’s focus on sandboxing: they open-sourced anthropic-experimental/sandbox-runtime (Apache‑2) and describe filesystem isolation plus network isolation implemented via a unix-domain-socket proxy that enforces domain allow-lists and handles user confirmations. On macOS/Linux they leverage seatbelt/Bubblewrap. That architecture aims to stop prompt-injection and “lethal trifecta” data-exfiltration attacks by cutting off network legs, enabling more permissive (YOLO-style) agent behavior while limiting risk. Caveats: the “trusted” default domain list is broad (potential exfiltration vectors) and cost/pricing for heavy use remains unclear, but the feature significantly raises the convenience and safety bar for hosted coding agents.