Fetching from the wire…
Public story · 2026-08-08 · high
Repos and secrets stay on your network, but prompts and results still leave for inference, and one user locks each runner.
Why now: The beta shipped in Claude Code 2.1.224 on August 7, one day after Anthropic's August 6 blog announcement.
Anthropic opened Claude Code's cloud sessions to self-hosted runners in public beta on August 6, per the Claude Code docs.
For Team and Enterprise accounts, sessions from the web, mobile, desktop, claude --cloud, or a scheduled routine can now run on hardware inside your network. That covers repo checkouts, build artifacts, and secrets, none of which touch Anthropic's infrastructure. The feature is off by default. It's excluded for Zero Data Retention orgs and for inference routed through Bedrock, Google Agent Platform, Microsoft Foundry, or an LLM gateway.
All traffic runs outbound only. Runners call api.anthropic.com over HTTPS, and Anthropic never connects inward, per the docs. Prompts, responses, and tool results still leave for inference, so code stays put but the conversation about the code doesn't.
The operational catch is in how runners get claimed. A runner locks to the first user who starts a session on it and serves only that account, up to a --capacity flag. Minimum fleet size equals your peak concurrent user count.
Runner lifecycle has two flags worth reading before a first deploy. --drain-grace-sec defaults to 0, so a runner exits immediately for a fresh disk. --retire-at <epoch-seconds> is required for kills you can't signal for, like a spot instance getting reclaimed.
A related change to Claude Code's sandbox now decodes JWTs and re-signs AWS SigV4, so agents can call authenticated services without seeing the secret credential. Self-hosted runners and masked credentials attack the same problem from opposite ends, keeping secret material away from the model.
Teams that provision this like a shared pool will under-provision. A runner claimed by one user won't take a second session from anyone else while it's warm. Fleet size has to track peak concurrency, not average load. Worth watching whether Anthropic loosens the one-user-per-runner lock as the beta matures.
The beta shipped in Claude Code 2.1.224 on August 7, a day after Anthropic's August 6 announcement.
Each link below shares sources, entities, or timing with this story.
Same source domain / Semantically similar
Reported by the same outlet (code.claude.com); covers closely related ground (similarity 0.77).
Same source
Cite the same source (Claude Code Docs — Self-hosted environments (announced claude.com/blog Aug 6, shipped in Claude Code 2.1.224 Aug 7)).
Same source domain / Semantically similar
Reported by the same outlet (code.claude.com); covers closely related ground (similarity 0.70).
Reported by the same outlet (code.claude.com); covers closely related ground (similarity 0.65).
Reported by the same outlet (code.claude.com); covers closely related ground (similarity 0.59).
Semantically similar
Covers closely related ground (similarity 0.81).
Covers closely related ground (similarity 0.80).
Covers closely related ground (similarity 0.77).