agenthold runs AI coding agents (Claude Code, Codex and others) on a server you control. Every agent gets its own Docker sandbox with a home directory that survives sleep and restarts, and people talk to it from a chat in the browser.
- One sandbox per agent. Agents sleep when idle and wake on demand. Files, history and sessions stay on disk.
- Secrets stay outside. An egress proxy injects credentials into outgoing requests, so agents use API keys and logins they never see. Egress and secrets explains the model.
- Shared and audited. Workspaces, roles, invitations and an audit log let a team use the same agents. See Sharing and audit.
- Scriptable. Everything the UI does is in the HTTP API, and tokens, schedules and webhooks drive it from scripts. Agents can call each other.
Try it at demo.agenthold.dev, or run your own.