What is Agent Autoloop?
Agent Autoloop turns a command-line coding agent into something that works without a person watching the terminal. A local hub receives events (a cron schedule, a GitHub or GitLab webhook, an HTTP call), matches them against rules in ~/.agent-autoloop/rules.json, and starts the named agent in the named folder. The agent finishes one turn and exits. If it needs a decision, it ends its last reply with a small JSON block, and the hub posts a card to a desktop notification, Feishu, DingTalk, WeCom, Slack or Discord. While the card waits, no agent process runs and no tokens are spent; once you answer, the hub resumes the same session.
Built-in agent plugins cover Claude Code, Codex, ACP agents such as Gemini CLI and OpenCode, DeepSeek Harness and ZCode, and any CLI that prints its last reply can be added with one line of config. Each rule sets an access level: read, edit (the default) or full. For Claude Code these map to the plan, acceptEdits and bypassPermissions permission modes; for Codex to read-only, workspace-write and danger-full-access sandboxes. The kernel stores every state change in an append-only SQLite log, so `agent-autoloop why <run>` explains any run. The main README is in Chinese, with an English version alongside.
Is Agent Autoloop worth it?
The useful idea here is that an agent waiting for your answer should cost nothing. Agent Autoloop ends each agent turn, stores the open question in its database and sends you a card; three days of silence means three days with no agent running. The repository was created on 3 November 2026 and had 5,076 stars, 924 forks and 207 commits by 13 November, when v1.8.3 shipped, its third release in four days.
The engineering is more careful than the age suggests. Events cannot choose the agent, the folder or the access level; only your local rules can, and an event that carries those fields is rejected. Issue titles and bodies from outsiders go into a separate file marked as material rather than into the prompt. Events an agent emits count as hops, and past 8 hops the hub stops dispatching and tells you. The maintainer reports a cold start of 88.78 ms and 59.94 MB of idle memory on an Apple M5 Pro, measured by the project itself.
What each run can do is set by the access field. The default, edit, lets Claude Code accept file edits without asking and gives Codex write access to the workspace. Full maps to bypassPermissions and danger-full-access, which means the agent runs any command with no approval. The project's own security notes say plainly that askFirst, the instruction to check with you before risky steps, is only a request in the prompt; the agent can ignore it and the hub cannot stop it. Generic command agents get no permission mapping at all.
The catch: this is a ten-day-old project, mostly the work of one maintainer though 30 accounts have contributed, with a Chinese-first README, an install that pipes a script from GitHub into your shell, and channel templates for Slack, Discord, DingTalk and WeCom that, by the README's own account, have only been tested against local mock servers. Card links work for whoever holds them for 7 days. The README also notes that providers prefer API keys over personal subscriptions for automated workloads, so running it on a Claude or ChatGPT plan may sit awkwardly with their terms.
Our take: for a morning brief or a nightly issue sweep on edit access it is a tidy upgrade over a cron job and a shell script; give a rule full access only where you would be comfortable with the agent running unattended as you.
Where it earns its place
A solo developer maintaining a small open-source library
Have new GitHub issues triaged and a local fix drafted before I sit down
The github-issues recipe polls every 5 minutes, hands each new issue to your agent on edit access, and notifies you to review the local change; it does not commit, push or reply on the issue. Cost is the agent usage per issue on your own account.
An on-call engineer at a ten-person startup
Get a first diagnosis of production alerts without waking up for every one
An HTTP source receives the alert, a rule sends it to Claude Code in the service folder, and any restart comes back as a card with buttons. Keep the rule on edit access so the agent cannot act on servers without your answer.
A product manager who lives in Feishu
Receive a weekly summary of reports and team tables in my own chat
The feishu-brief recipe runs on Sunday at 18:00 Beijing time, reads with your own Feishu identity, keeps the full draft on your machine and sends only a short summary to your private chat.
Before you commit
- Setup
- Needs Node 24 or newer, Git and an agent CLI you have already signed into. Install with the PowerShell or curl script, then run `agent-autoloop try` for a full ask, answer and resume cycle.
- Integrations
- Claude Code, Codex, ACP agents (Gemini CLI, OpenCode), DeepSeek Harness, ZCode and generic CLIs; GitHub and GitLab webhooks, HTTP and cron sources; desktop, Feishu, DingTalk, WeCom, Slack and Discord channels; an MCP server for reading and answering cards.
- Data and privacy
- The hub listens on localhost only and keeps its database, event text and run output in a data directory set to 0700. It launches your signed-in CLIs without reading their credentials; whatever the agent reads goes to its model provider as usual.
- Cost at scale
- The software is free; each run is billed to the agent account it uses. The README points to API keys rather than personal subscriptions for heavy automation.
- Limits
- Access levels map to CLI permissions only for the built-in plugins; generic commands are not restricted. History growth slows status summaries, and a format change rebuilds state from the log on first start.
What’s good
- Waiting for a decision costs no tokens and no running process
- Events cannot set the agent, folder or access level; only local rules can
- Append-only log explains every run with `agent-autoloop why`
- Plugins for Claude Code, Codex, ACP agents and any CLI
What’s not
- Created on 3 November 2026; releases land every day or two
- askFirst is a prompt instruction the hub cannot enforce
- Slack, Discord, DingTalk and WeCom templates tested only on local mocks
- Main README and most docs are in Chinese
Best for
Developers who already run Claude Code or Codex and want scheduled or webhook-driven jobs to run unattended, with a phone notification for the decisions.
Agent Autoloop alternatives
Agent Autoloop: common questions
- Is Agent Autoloop worth it?
- Agent Autoloop is a carefully built way to let Claude Code or Codex work through a queue while you are away, but it is ten days old and its safety rests on the access level you choose, not on the hub. We rate it 3.8 out of 5.
- How much does Agent Autoloop cost?
- Free and open source (MIT). Model usage is billed to the Claude Code, Codex or other agent account you are already signed into.
- What are the downsides of Agent Autoloop?
- Created on 3 November 2026; releases land every day or two; askFirst is a prompt instruction the hub cannot enforce; slack, Discord, DingTalk and WeCom templates tested only on local mocks; main README and most docs are in Chinese.
- Who is Agent Autoloop best for?
- Developers who already run Claude Code or Codex and want scheduled or webhook-driven jobs to run unattended, with a phone notification for the decisions.
- How much does Agent Autoloop cost for a team or in production?
- The software is free; each run is billed to the agent account it uses. The README points to API keys rather than personal subscriptions for heavy automation.
- Is Agent Autoloop safe for company data?
- The hub listens on localhost only and keeps its database, event text and run output in a data directory set to 0700. It launches your signed-in CLIs without reading their credentials; whatever the agent reads goes to its model provider as usual.
- What are the best Agent Autoloop alternatives?
- The closest alternatives we have reviewed are Claude Squad, Vibe Kanban and n8n.
Answers checked Nov 13, 2026.