Self-Hosted Claude Code Runs on Your Boxes. The Conversation Still Leaves Your Network.
Self-hosted environments went public beta on August 6. Checkouts and secrets stay on your infrastructure — prompts, responses, and tool results still go to api.anthropic.com. And a runner serves one user at a time, so your minimum fleet size is peak concurrent developers.
NeuroX AI · August 12, 2026
Anthropic put self-hosted environments for Claude Code into public beta on August 6, for Team and Enterprise plans. Cloud sessions — web, mobile, desktop, scheduled routines, claude --cloud — execute on runners inside your network instead of Anthropic's. The docs state the boundary plainly: "The conversation itself, including prompts, responses, and tool results, goes to api.anthropic.com for model inference." Repository checkouts, build artifacts, and secrets stay on machines you provision. The transcript is stored so a session resumes from another surface.
Then there's the sizing math nobody puts on a slide. A runner locks to the account of the first user whose session it claims and then serves only that user, up to its configured --capacity. That's how checked-out code stays unmixed without wiping disk between users — and it means your minimum fleet size is the number of developers active at once, not the number of sessions.
The operational edges are where pilots die. Stop polling for about 60 seconds and the control plane requeues the session to a different runner. --drain-grace-sec defaults to 0, so a runner exits as soon as its sessions finish and your orchestrator restarts it with a fresh disk. On spot instances, pass --retire-at — a signal-less host kill is indistinguishable from a crash.
Read the exclusions before you plan a rollout: unavailable with Zero Data Retention, and inference can't route through Bedrock, Google, Foundry, or an LLM gateway. Self-hosting moves execution, not the control plane.