Agent Autoloop: agents keep working after you close the terminal
Your agents pick up work in the background and send you a card only when the call is yours
- Who
- Made by KKKKhazix for people who use command-line agents such as Claude Code and Codex. MIT licensed.
- When
- The repository opened on 3 November 2026; version 1.8.3 came out on 13 November.
- Where
- Your own PC on Windows, macOS or Linux. Works with Claude Code, Codex, agents that speak ACP and other command-line agents.
What was hard
Most of us use Claude Code or Codex by sitting at a terminal: type a prompt, watch it run, answer when it asks. Close the window or walk away, and the work stopped there.
What it does about it
Agent Autoloop is a small hub you install on your own PC. When it is 9:00 AM or a new GitHub issue arrives, it hands the work to an agent by the rules in rules.json. The agent works one round and exits; for a decision that is yours, such as deleting data or restarting a service, it sends you one card. While it waits no agent process runs and no tokens are spent, and your answer picks up the original conversation to finish the job.

What people made with it
Ran one Codex session overnight and tabled the input tokens read each round: while it waited for an outside API and kept resuming the same session, even an idle round cost 180–190K tokens.
实测缓存命中率始终在 97% 左右,因此主要开销并非来自缓存过期,而是长会话本身的体量累积。
Connected the Feishu channel to their own app: cards were never sent on lark-cli 1.0.66, and the same setup worked after moving to 1.0.100.
2026-11-12 真机实测(本机先在 1.0.66 上复现,换 1.0.100 后同一配置即正常)
How it runs inside
- 1
Events (cron, GitHub) → The hub's kernel
One standard eventA webhook or a schedule becomes an event with a type, a subject and a link.plugins/github/source.ts - 2
The hub's kernel → worker
A job, by the rules in rules.jsonOnly the rules on your PC decide which agent runs and in which folder.kernel/decide.ts - 3
worker → Claude Code / Codex
claude -p … --session-idThe agent runs detached; what people wrote goes in a context file as data, not in the prompt.plugins/claude/agent.ts - 4
Claude Code / Codex → worker
Final message + agent-autoloop blockIt writes what it needs of ask, notice, resolve, later and events, and the process ends.worker/contract.ts - 5
The hub's kernel → You
One cardA desktop notification by default; Feishu, DingTalk, WeCom, Slack and Discord too.plugins/desktop/channel.ts - 6
You → The hub's kernel
agent-autoloop answer 1 2You answer with the card's number and the option's number.cli/main.ts - 7
worker → Claude Code / Codex
claude -p … --resumeThe session that asked is carried on and the rest of the work is done.plugins/claude/agent.ts - 8
Claude Code / Codex → Finished work, in status
A one-line conclusionagent-autoloop status keeps it with rounds, tokens and cost.cli/status.ts
What goes in, what comes out
Is it for you
Good for
- Getting a brief or an inbox triage at the same time every day
- Having an agent take a first look at new issues or production alerts
- Labelling a PR and leaving review through deploy check to the agent
- Running agents in the background on one PC, with no server
Not for
- CLI output and default notification strings are still Chinese (per the English README).
- Resuming a long session costs tokens even on an idle round: one user measured 180–190K per round on Codex (issue #268).
- For high-frequency automation the README points to the vendors' advice to use an API key rather than a subscription.
- The DingTalk, WeCom, Slack and Discord templates were checked only against a local mock server.
Reviewed by Devalley’s editor before publishing. How the column is made