Your CLI agent, finally allowed to work unsupervised
A tiny event-sourced kernel that lets Claude Code or Codex take jobs, pause for your approval, and resume — without burning tokens while it waits.
Agent Autoloop is a 1,651-line, zero-dependency kernel that lets command-line coding agents take background work end-to-end — and burn nothing while waiting for your decision.
What it does
Agent Autoloop is a self-hosted automation hub for command-line agents like Claude Code and Codex. Events (cron jobs, GitHub issues, alerts) flow in, rules decide which agent handles what in which directory, and the agent runs to completion. When it hits something risky — deleting data, restarting a service, spending money — it pushes a card to Feishu, Slack, Discord, or desktop notifications and waits for your click before continuing.
The interesting bit
The wait is free. When an agent needs your decision, its process exits entirely; the pending state lives in a database, so no tokens are burned while you’re at lunch — or on a three-day vacation. When you answer, it resumes the original conversation across processes. The kernel itself is ~1650 lines with zero runtime dependencies, deterministic (no clock reads, no randomness, no outbound requests), and event-sourced: all state derives from an append-only log, so no migration scripts and every decision is auditable.
Key highlights
- Any CLI tool that accepts a prompt and prints its last reply can be plugged in — no SDK, no agent source changes
- Agents signal via a simple
agent-autoloopcode block in their final reply: ask, notify, resolve, schedule a recheck, or trigger new events - Malformed agent JSON triggers a rewrite of just that reply, not a redo of the work already done
- Ships with ready-made recipes: morning briefs, GitHub issue triage, ops alert investigation, website monitoring, PR delivery (review → CI → merge → deploy → verify)
- Runs on your own Mac/Linux box — no cloud server, no open ports
Caveats
- Several webhook channels (DingTalk, WeCom, Slack, Discord) were only tested against a local mock service, not real accounts; same for the sample recipes’ fake data
- The “other agents” integration commands (DeepSeek, Cursor, pi, etc.) were compiled from official docs but not individually tested
- Power-loss durability is explicitly untested; process kills are covered
Verdict
Worth a look if you already live in Claude Code or Codex and want them doing scheduled or event-driven work without a babysitter. If you need a hosted, multi-user, web-UI orchestration platform, this is deliberately not that.
Frequently asked
- What is KKKKhazix/agent-autoloop?
- A tiny event-sourced kernel that lets Claude Code or Codex take jobs, pause for your approval, and resume — without burning tokens while it waits.
- Is agent-autoloop open source?
- Yes — KKKKhazix/agent-autoloop is open source, released under the MIT license.
- What language is agent-autoloop written in?
- KKKKhazix/agent-autoloop is primarily written in TypeScript.
- How popular is agent-autoloop?
- KKKKhazix/agent-autoloop has 4.1k stars on GitHub.
- Where can I find agent-autoloop?
- KKKKhazix/agent-autoloop is on GitHub at https://github.com/KKKKhazix/agent-autoloop.