Agent AutoloopPartly tested🇨🇳 CN

Khazix's Agent Autoloop dispatches CLI agents in the background, exits when they are waiting, and only sends a Feishu or desktop card when it needs your decision

With Claude Code or Codex, work usually stops the moment you close the terminal or walk away. Chinese-community author 数字生命卡兹克 / Digital Life Khazix (@Khazix0918, GitHub KKKKhazix) open-sourced the dispatcher he had used for a year as Agent Autoloop (MIT): a local, zero-runtime-dependency kernel that reads rules.json and hands cron jobs, GitHub issues, alerts and other events to a CLI agent you already signed into. The agent runs one round and exits. Deleting data, rebooting, spending money or messaging the outside world becomes a card on Feishu, DingTalk, WeCom, Slack, Discord or the desktop; while you decide, nothing sits around burning tokens, and answering resumes the same session. The kernel claims to do only four things—ingest events, route them, dispatch an agent, and record what happened—with zero-cost suspend, event sourcing, separating instructions from untrusted input, and a hop circuit breaker at 8 layers. Events, agents, channels and delivery are plugins. Bundled recipes cover a morning brief, GitHub issues, alerts, website watch, Feishu weekly notes, inbox triage, feedback and PR delivery. Install by handing the repo to your agent (it is told to read INSTALL.md), or run curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh on macOS/Linux (irm .../install.ps1 | iex on Windows). Node 24+ is required. Windows support landed on 6 November. We cloned main on Linux (fade2ce, after the v1.6.4 release) and ran npm run test:quick with Node 24.11.0: 697 of 719 tests passed, 0 failed, 22 skipped because this machine is not Windows/macOS or has no Claude Code. We did not run the installer, agent-autoloop try against a live Claude/Codex login, or Feishu.

Visit site Agent Autoloop ↗
Pricing
Free
Region
Global
Where it's from
Chinese community
Added
Nov 9, 2026
Tested on
Nov 9, 2026
Version we used
agent-autoloop main @ fade2ce (after release v1.6.4); Node 24.11.0 (npm run test:quick)
Categories
Automation & AgentsCodingProductivity
Original author
数字生命卡兹克 @Khazix0918 ↗
Our test
Partly tested
Cost / time
The repo is MIT/free and the published package has no runtime dependencies. You need Node 24+ on the machine and a signed-in CLI agent (Claude Code, Codex, etc.) whose subscription or quota you already pay for. try, plus every real run, retry and resume, spends model tokens; waiting for your answer does not. Feishu, GitHub issue polling, mail triage and similar channels use those services' own accounts. Autoloop itself does not add a fee.
Source
X @Khazix0918 ↗X @Khazix0918 ↗X @Khazix0918 ↗X @paiven_x ↗GitHub KKKKhazix/agent-autoloop ↗

The task

Let the Claude Code or Codex session you already use pick up cron jobs, issues and alerts in the background, so you do not have to watch the terminal—and so it only interrupts you for a real decision, then resumes the same conversation.

Steps

  1. You need Node 24 or newer, Git, and a CLI agent you have already signed into (Claude Code, Codex, or any command that takes a prompt and prints a last message). macOS/Linux also needs curl. No GitHub account is required.
  2. Hand https://github.com/KKKKhazix/agent-autoloop to your agent and tell it to read INSTALL.md and the README, then install and run agent-autoloop try. Or install yourself: curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh on macOS/Linux, or irm https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.ps1 | iex in Windows PowerShell. install and update pull the latest checked Release.
  3. Verify with agent-autoloop --version, agent-autoloop doctor and agent-autoloop try. try spends model quota on a full ask → exit-and-wait → you answer → it finishes loop. The terminal prints a card id; type the option number in an interactive terminal, or run agent-autoloop answer <card> <option> if it is not interactive. Use the number on your screen, not a copied 1. Then run agent-autoloop status and look for a successful trial. If try prints the question and exits in a non-interactive terminal, that is the wait, not a crash.
  4. Desktop notifications are on by default, so a local trial does not need Feishu. For phone cards: npm i -g @larksuite/cli, then lark-cli config init --new, enable the bot and card callbacks (long connection) in the developer console. Create a channel key with agent-autoloop key new channel feishu and put it, plus your open_id, in config.json; list yourself in rules.json channels.feishu.answerers. That key is Autoloop's channel secret, not the Feishu app secret. Only one listener can hold an app's card clicks, so run it on one machine.
  5. Edit ~/.agent-autoloop/rules.json: what happens, who does it, where, and how. Example morning brief: {"name":"Morning brief","when":{"cron":"0 9 * * *"},"to":"codex","cwd":"~/notes","prompt":"Write today's brief."}. to must match a name already registered under agents; timezone is a top-level field (for example Asia/Shanghai), not a field on the rule. Set "tell":"always" if you want a notice every round; the default notifies only when there is a card. Validate with agent-autoloop check, or check a draft file first. Changes apply in seconds. agent-autoloop rules explain shows the next run; agent-autoloop run "rule name" runs it now.
  6. Or start from a bundled recipe (morning brief, GitHub issues, alerts, website watch, Feishu weekly notes, inbox triage, feedback, or delivering a PR tagged deliver): copy its task skill into the agent, merge the sample rule, replace hostnames and paths, then check. The repo also ships an agent-autoloop skill so your agent can add rules, plugins or machines for you.

Prompt

Install this Agent automation hub: https://github.com/KKKKhazix/agent-autoloop
Read INSTALL.md and the README first, then install it and run agent-autoloop try to verify. Tell me the install path, which agent it targeted, and the verification result.

What the author says

Author @Khazix0918 (open-source essay on 4 Nov 2026; Windows support announced 6 Nov): the kernel is about 1,600 lines with no third-party deps; his own stress test started in under 0.1s, used about 60MB idle and about 170MB with 1,000 connections, and dispatched work in under 1.5ms. While you decide, the process exits and spends no tokens even if you take three days. External issue text is stored as data, not mixed into instructions. Agent-to-agent wakeups stop after 8 hops and notify you. He listed jobs he actually runs: an afternoon roundup of how people talk about AIHOT, a weekly browser GEO check, an evening sweep of free quotas, a morning price check, a five-minute uptime probe that costs nothing when the site is up, and a Sunday cleanup of duplicated code. Inside his company's internal workbench, colleagues message a Feishu bot; Claude evaluates, patches and ships—about 70 of 80+ notes handled on its own, a dozen decided by him, often within an hour when automation can do the work. The current English README now says about 1,850 kernel lines and publishes his benchmarks (88.78ms cold start, 1.82ms dispatch p99). @paiven_x's 8 Nov Windows 10 write-up used 22H2, release v1.6.3 and Node 24.21.0: after install and Feishu, a background Codex run produced a sourced research note, then a multi-agent collaboration playbook. The stalls he hit were a network fetch that could not read public pages, and a Feishu multi-agent layer he added himself that could not reassign work after he approved a fix—he said that gap was not Autoloop's.

Our test

Cloned KKKKhazix/agent-autoloop main at commit fade2ce (9 Nov 2026, 03:09 +08, after the v1.6.4 release on 8 Nov). Read the bilingual README, INSTALL.md and skills/agent-autoloop/SKILL.md, and counted TypeScript lines under kernel/. Ran npm run test:quick on Linux with a local Node 24.11.0. Did not run install.sh or install.ps1, did not install the login service, and did not run agent-autoloop try against Feishu or a signed-in Claude Code/Codex.

What we found when we tried it:test:quick reported 719 tests, 697 passed, 0 failed, 22 skipped, about 55.8 seconds, exit code 0. The skips were mostly Windows login-task/PowerShell/short-path cases, macOS osascript, and a live Claude Code adapter check. kernel/ has 11 TypeScript files totalling 1,930 lines including blanks and comments; the original post said about 1,600, and README.en now says about 1,850. package.json has no dependencies, engines.node is >=24, and the license is MIT. examples/recipes/ contains eight templates: browser-handoff, feedback, feishu-brief, github-issues, inbox-triage, morning-brief, ops-alerts and website-watch.

How to get in

Public GitHub, MIT, with English and Chinese READMEs. Runs on Windows, macOS and Linux; needs Node 24+ and a signed-in CLI agent. Feishu, DingTalk and other channels need their own CLIs. The original essay is in Chinese; @paiven_x published a separate Windows 10 practice note.

Who it suits

People who already run Claude Code or Codex and want cron jobs, issues and alerts followed through in the background, interrupting them only for decisions. Skip if you have no local CLI agent, or do not want a long-running hub on the machine.

Pitfalls

  • Node 24 or newer is required; 20/22 is not enough. Windows support landed on 6 November 2026—older posts still said it would arrive that weekend.
  • agent-autoloop try needs a signed-in agent and spends quota. Card ids change; if a non-interactive try prints a question and exits, it is waiting, not failing.
  • timezone belongs at the top of rules.json, not on a single rule; to must be a registered agent name; set "tell":"always" if you want a notice every round.
  • The Feishu key in config.json is Autoloop's channel secret, not the Feishu app secret. Only one machine can listen for that app's card clicks.
  • Waiting is free; running, retrying and restoring context are not. The README says some webhook and Feishu templates were checked against local mocks or a fake CLI, not a live account end to end.
  • We only ran the repo's test:quick suite. We did not execute the official installer or walk try through a live Claude/Codex session or Feishu.

More in this category