Claude Code sandbox now gates network per command, not per session

Anthropic's dev platform shipped a batch: fast mode on remote runners, scored plugin evals, and on-demand API compaction land alongside a tighter sandbox.

Nowline SEP 16 4:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The allowlist travels with the command

    In sandboxed auto mode, the hosts a Bash, PowerShell, or Monitor command needs are now reviewed alongside that command and opened for it alone — every other host is refused. Unattended and auto-approved agent runs get network egress scoped per step, not one blanket session grant.

  • Fast mode reaches Remote sessions

    Claude Code Remote — cloud and self-hosted runners — now honors fast mode: flip the host setting or type `/fast` where your org allows it. Edge cases got cleaned up too, so `/fast off` actually turns it off and fast requests fall back to standard speed on overloads or credit limits instead of failing the turn.

  • `claude plugin eval` scores your plugins

    Plugin authors can run a plugin's eval suite against Claude Code and get scored, reproducible results as JSON plus an HTML report (`claude plugin eval --help`). It turns "seems to work" into a regression check you can drop straight into CI.

  • On the API: compaction when you say so

    The Messages API's on-demand compaction (beta header `compact-2026-09-04`) takes `compaction: {type: "summarize"}` and returns a signed summary block, with keep-tail so recent turns stay verbatim. Long-running agents can compress context on your schedule rather than only at a fixed token threshold.