Agent Autoloop is a new self-hosted automation hub for people who already run command-line coding agents and want them to keep working when the terminal is closed. The KKKKhazix/agent-autoloop repository went public on November 3, 2026 under the MIT licence and quickly drew thousands of GitHub stars. The author describes about a year of private use before the open rewrite.
The kernel is intentionally small. Documentation puts it near 1,650 lines with no runtime dependencies beyond what the install script expects on macOS or Linux: Node 24, Git, curl and a logged-in agent CLI. Windows support is explicitly deferred. Work arrives as events from cron rules, GitHub or GitLab webhooks, HTTP callbacks or issue polling. Rules decide which agent runs in which directory. The agent completes one round, writes its result into its final reply and exits so no process sits idle while waiting for a human.
When the agent needs a consequential choice, such as deleting data, restarting a service, spending money or messaging outside, the hub sends an approval card through desktop notifications or channels including Feishu, DingTalk, WeCom, Slack and Discord. After the user answers, the hub resumes the same session. The project frames four plugin directions: event sources, agents, notification channels and a delivery pipeline that can shepherd a pull request through review, CI, merge and release checks.
Built-in agent adapters cover Claude Code, Codex and ACP, plus a generic command plugin for other CLIs. Scenario templates include morning reports, GitHub issues, alerts, feedback triage and website checks. A SECURITY.md warns that rules control permissions, that prompt injection from external issue text is not solved, that read access can still reach secrets on the machine, and that approval links act as bearer tokens for a limited time.
Independent write-ups on 人人都是产品经理 and explainx.ai described the release within a day of the public debut, focusing on the background loop and the approval-card design. Releases moved quickly from v1.0.0 on November 3 to v1.2.0 on November 5 with install-from-release updates and session handoff fixes. The hub does not replace model vendors; it schedules the tools you already pay for and keeps risky steps behind an explicit human gate.