loopx

Overview

Agent-agnostic local control plane that keeps durable goals, gates, todos, evidence, and quota stable across Codex, Claude Code, and other coding agents. It provides executable todos, quota-aware auto-wake, evidence logs, and verifiable handoffs to keep long-running work reviewable and restartable.

README Preview

\n\nLoopX\n\n\n\n**The local control plane for long-running AI agent work.**\n\nKeep objectives, gates, todos, evidence, quota, and handoffs stable while Codex, Claude Code, Cursor, or your own runtime executes bounded turns.\n\n[](LICENSE) [](https://github.com/huangruiteng/loopx/releases/latest) [](pyproject.toml) [](docs/public-private-boundary.md) [](docs/product/release-readiness.md)\n\n[Public website](https://huangruiteng.github.io/loopx/) · [Docs](https://huangruiteng.github.io/loopx/docs/) · [Try LoopX](#try-loopx) · [See real loops](#evidence) · [How it works](#why-loopx) · [User manual](https://my.feishu.cn/wiki/CaL5wMk9ui17ngkWzeUcMlAYnZg) · [简体中文](README.zh-CN.md)\n\n**把会干活的 Agent，接成可管理、可复盘、可持续改进的数字员工。**\n\n\n\n---\n\nA lightweight state kernel and agent-agnostic local control plane for loop\nengineering, LoopX keeps long-running work reviewable, restartable, and easier\nto hand off across turns, tools, and agents. It does not replace your agent\nruntime.\n\n**Loop engineering for long-running AI agents and peer agent teams.**\n\n> Keep the loop moving. Keep the judgment human.\n\n\n\n## Why LoopX\n\nAn agent can finish a task in one session. Long-running work is harder:\nobjectives change, owner decisions appear, evidence goes stale, agents hand work\nto peers, and a scheduler can keep spending after no useful transition remains.\nChat memory and a timer are not enough to govern that.\n\nLoopX keeps the durable control state in one compact layer:\n\n```text\nobjective / issue / project\n   │\n   ▼\nLoopX state: objective + gates + todos + scope + evidence + quota\n   │\n   ├─ human judgment needed? ── yes ─▶ ask a concrete question and wait\n   │\n   ├─ safe fallback available? ──────▶ run one bounded agent slice\n   │\n   ▼\nCodex / Claude Code / Cursor / shell agent executes one turn\n   │\n   ▼\nwrite evidence + handoff + next todo ─▶ quota decides the next tick\n```\n\n\n\nA useful mental model is an\n**[agent-native Kanban for long-running work](d