Analysis
Anthropic opened a public beta this week for self-hosted environments in Claude Code, letting sessions run on infrastructure a customer's own team controls rather than on Anthropic's cloud, the company said in a blog post. Sessions can still be started from the web, mobile, desktop app or a scheduled routine, but execution -- repository checkouts, build artifacts, secrets, and any files a session creates -- now stays on servers the organization provisions inside its own network, next to its internal services and security controls.
The feature targets a specific gap: teams whose compliance, data-residency or internal-tooling requirements meant they couldn't put Claude Code sessions on Anthropic-hosted infrastructure at all, even with existing permission and sandboxing controls. It's available in public beta to organizations on Claude's Team and Enterprise plans, off by default, and explicitly not available to organizations using Zero Data Retention, according to Unite.AI. Anthropic says organizations need a platform or developer-productivity team to own setup -- building and maintaining the runner image, keeping runners updated, and running the orchestrator for on-demand mode -- which puts real operational weight on the customer, not just Anthropic.
The move follows a run of enterprise-focused Claude Code releases this year and lands the same week AISI's cybersecurity-testing disclosures raised fresh questions about how well sandboxed environments actually contain agentic coding tools. Self-hosted execution doesn't resolve that containment question on its own -- it shifts responsibility for the sandbox to the customer's own infrastructure team, rather than solving the underlying isolation problem.
