KKKKhazix/agent-autoloop

What it does

Agent Autoloop is a lightweight automation hub for command-line coding agents such as Claude Code and Codex. Instead of watching a terminal session, you register rules that describe what event, which agent, and which directory; the kernel then schedules work, runs the agent for one round, records the result, and exits the process. When the agent needs a decision (for example, deleting data, restarting a service, spending money, or sending a message), it emits a card to desktop notifications or Feishu, and the agent resumes the same conversation after you answer. During that wait no agent process is running and no tokens are spent.

The project splits its design into four plugin directions — event sources, agents, notification channels, and delivery pipelines — so each part can be replaced. The built-in kernel is described as about 1,650 lines with no runtime dependencies, and it uses an append-only event log, instruction/data separation, and a hop-count circuit breaker as its main safeguards architecture and mechanisms.

It targets individual developers and small teams who already have a logged-in CLI agent and want it to pick up routine work (scheduled reports, GitHub issues, alert triage, website checks) without keeping a terminal open.

Quick start

Requirements: macOS or Linux, Node 24, Git, curl, and any command-line agent you have already installed and logged in to. No GitHub account is required.

1. Install with the official script:

curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh

2. Run the built-in trial, which exercises the full ask → exit → answer → continue cycle:

agent-autoloop try

3. When the agent asks a question, answer it in an interactive terminal by typing the number, or in a non-interactive shell with:

agent-autoloop answer 1 2

The example chooses option 2 (“green”). Then check progress and usage:

agent-autoloop status

The status output shows the resumed conversation and today's and last-7-day rounds, tokens, and cost; the counting rules are described in the skill documentation.

4. For real work, edit ~/.agent-autoloop/rules.json. A rule names the event, the agent, and the working directory:

{
  "name": "Morning brief",
  "when": { "cron": "0 9 * * *" },
  "to": "codex",
  "cwd": "~/notes",
  "prompt": "Summarize yesterday's notes and commits into today's brief."
}

Changes take effect within seconds; run agent-autoloop check to validate them. Invalid edits keep the old rules running and produce a warning card.

Ready-made recipes include daily briefs, GitHub issue follow-up, alert triage, user feedback handling, website monitoring, and Feishu weekly summaries; each comes with rules and a task skill recipes.

Constraints and caveats

• Platform: macOS and Linux only; no Windows support is stated.
• Credentials: the hub calls your locally installed CLI tools and does not read or store their credentials; model usage is billed to your own account under each vendor's terms. For high-frequency automation or serving external users, the README recommends API keys rather than personal subscription quotas.
• Channel verification: Claude Code, Codex, and ACP were reportedly tested for multi-round Q&A and cross-process session resumption; Feishu was tested for sending, button callbacks, and card updates. The four webhook templates (DingTalk, WeCom, Slack, Discord) were only validated against a local mock service. Text channels rely on answer links, so hub.publicUrl must be reachable from your phone or computer.
• Other agents: one-line use: "command" integrations exist for DeepSeek Harness, pi, OpenClaw, Hermes, Cursor, and CodeBuddy, but the README says these commands were collected from vendor docs and have not been individually run.
• Sample data: the feedback, website-watch, and Feishu recipes use fictional sample data and were only validated with local HTTP and a fake Feishu CLI, not against real accounts.
• History growth: status aggregation slows as history grows, and the first startup after a format change rebuilds state from the full event log (43.1 seconds for 1.1 million facts).
• License: MIT.

HotRepo's take

On the HotRepo new-project board for 2026-11-04, the repository was recorded at 429 stars and 87 forks, with 429 stars gained during the measured window that began 2026-11-03. That is a single snapshot rather than a trend line, so it supports only the fact that the project attracted early attention on its first day.

The design choices behind that attention are concrete: a zero-dependency kernel with an append-only log, agent sessions that exit while waiting for approval, and instruction/data separation that keeps external Issue text out of the prompt. The README is also unusually explicit about what was not tested — the webhook channels, the additional agent commands, and the sample recipe flows — which matters more than a list of star counts for anyone deciding whether to run this against real repositories or accounts.