Claude Code 2.1.271 brings fast mode to remote sessions and per-command network allowlists

Anthropic released Claude Code 2.1.271 on 15 September 2026. Two changes stand out for web teams: fast mode now reaches remote sessions running in the cloud or on self-hosted runners, and a per-command network allowlist lands in the sandboxed tools.

#Fast mode reaches remote sessions

Claude Code’s fast mode runs on the Opus model with accelerated output; it does not fall back to a lighter model. Previously limited to local sessions, it now covers remote sessions — the ones where the agent works not on the developer’s machine but inside an ephemeral container, in Anthropic’s cloud or on a self-hosted runner. An organization-level setting turns it on for a whole team.

The payoff is real for shops already running agents at scale: the iteration loop gets shorter without giving up isolation. The repository is cloned fresh when the container starts and reclaimed after a period of inactivity, which limits unmanaged state leaking between runs.

#A per-command network allowlist

The second change targets execution safety. An agent that runs shell commands can, in principle, reach any address. Version 2.1.271 introduces a per-command domain allowlist in the sandboxed tools — Bash, PowerShell and the Monitor tool. Instead of allowing or denying the network wholesale, the operator declares which domains each command may contact.

The mechanism extends the permission model already exposed in settings.json, where allow and deny rules scope the tools. The example below illustrates the principle — the exact syntax for network scoping is documented in the changelog:

{
  "permissions": {
    "allow": [
      "Bash(npm test:*)",
      "WebFetch(domain:registry.npmjs.org)",
      "WebFetch(domain:api.github.com)"
    ],
    "deny": [
      "WebFetch(domain:*)"
    ]
  }
}

The logic mirrors how agents connect to their tools: narrowing what an agent can reach matters as much as choosing what it can do.

Scoping the network per command shifts the question from “should we trust the agent” to “what can it reach”.

#The rest of the release train

2.1.271 also brings quieter additions: support for omitClaudeMd in subagent frontmatter, to run them without loading local CLAUDE.md files; mouse support in the fullscreen /config panel; and a hint pointing Bedrock, Vertex AI, Foundry and LLM-gateway users to the desktop app. It sits in a tight sequence:

#The bottom line

Version 2.1.271 changes neither the model nor the agent’s core behaviour. It consolidates two trends of the year: running coding agents off the workstation, in isolated containers, and finely controlling what those agents can reach. For a team that hesitated to hand client code to an agent, per-command network scoping is a stronger argument than raw speed.

Claude Code’s very tight release cadence — three versions in five days — remains its signature: changes arrive in small strokes, best followed through the changelog rather than quarterly announcements.

Further reading: the Claude Code changelog (CHANGELOG).