Claude Code's New Self-Hosted Mode Keeps Source Code On-Premises. Prompts Still Leave the Building.

Anthropic launched self-hosted environments for Claude Code, now in public beta for Team and Enterprise subscribers, letting organizations run Claude Code sessions on their own infrastructure instead of Anthropic-hosted systems. The architecture runs on “runners” — long-lived processes executing sessions in isolated checkouts — in either a fixed mode with a set number of always-on instances or an on-demand mode that spins runners up as sessions queue and shuts them down once work clears. The pitch is network isolation (sessions reach internal databases without exposing them publicly), customization (pre-installed compilers and internal tooling), and data control (source code and build artifacts stay on customer infrastructure). One detail cuts against the security framing: conversations and prompts still route to Anthropic for inference, and the feature is explicitly not available to organizations using Zero Data Retention. This ships for teams with strict network requirements — not the ones with the strictest data requirements.

Anthropic’s own engineering org is a useful reference point for what “self-hosted, governed Claude Code” looks like in practice. Deputy CISO Jason Clinton has detailed the five practices securing a pipeline where roughly 80% of merged code is now AI-authored: automated security reviews benchmarked against MITRE ATT&CK, CLAUDE.md files that encode secure-coding standards directly into every session’s context, PR-review agents dedicated solely to security issues, continuous dynamic testing in staging, and triage agents running under strictly limited permissions. Clinton says internal analysis found roughly a third of past incident-causing bugs would have been caught earlier had these controls existed at the time.

The scale case for infrastructure-level control shows up in Anthropic’s Alberta government case study: a small team scanned 466 million lines of code across 3,400 repositories in about 20 hours — roughly 6.5 years of estimated manual effort — running some 50 agents in parallel, with every generated fix reviewed and approved by the ministry’s own engineers before shipping. That’s the shape of workload self-hosted runners are built for: massive, parallel, and accountable to a human at the point of every change.

For teams evaluating Claude Code at enterprise scale, the self-hosted beta answers “where does my code live” without yet answering “where does my prompt go” — worth reading the fine print before treating this as a full data-residency solution.