loopx

LoopX is an open, local-first control plane for long-horizon agents and personal agent teams. It keeps goals, decisions, evidence, todos, scope, and quotas durable across sessions, restarts, and different agent runtimes. The project positions itself as an agent-native control layer rather than another agent framework: runtimes such as Codex, Claude Code, Cursor, DeepSeek Harness, KunlunCode, OpenCode, Pi, ZCode, and custom runners execute the work, while LoopX governs continuation, ownership, gates, and handoffs. The Personal Agent Workspace provides a browser/PWA dashboard (`loopx dashboard`) for viewing what needs attention, running tasks, scheduled watches, capability overrides, deliverable files, and supporting evidence. It supports steering live turns, queuing messages, and routing async inbox messages from connected Lark conversations. LoopX also includes a command-line interface with operations such as `quota should-run`, `todo claim`, `todo update`, `refresh-state`, and `quota spend-slot` for custom runner integration. The README reports benchmark results on the Long-Horizon Terminal-Bench (LHTB): LoopX 1.0.3 Heartbeat reached 0.4948 mean reward across 46 matched tasks using GPT-5.6 Sol, compared with 0.4218 for Plain Codex and 0.4475 for native Codex Goal. The authors note that strict solve counts matched Plain Codex, that task-level results improved on 23 tasks, tied on 13, and regressed on 10, and that the study uses one effective trial per task and mode with unequal runtime budgets. Additional exploratory studies cover SWE-Marathon and DeepSWE behavior analysis. The project also presents long-running showcase cases, including a 200+ hour OpenViking contribution sequence and a redacted Auto ML experiment trajectory, plus user-reported cases such as a 13+ hour C++ accuracy run, a four-day unattended run, and seven merged PRs. Installation is via PyPI with Python 3.11+ and Node.js 22.22.3+ (Node.js 24 LTS recommended). Basic usage statistics default on after first-use disclosure and can be disabled. The project is Apache-2.0 licensed and provides documentation, reproducible workspace scenarios, and a public showcase catalog.