LoopX is a provider-neutral state kernel and local-first control plane for what its author calls loop engineering — keeping long-running agent work reviewable, restartable, and handoff-safe across many bounded turns. Its central claim is a separation of concerns: LoopX governs durable state, while Codex, Claude Code, Cursor, or a custom runner does the actual work. It explicitly does not replace the agent runtime and is not another orchestration framework.

The problem it targets is the one that appears after the single-session demo ends. Objectives drift, owner decisions pile up, evidence goes stale, agents hand work to peers, and a scheduler happily keeps burning quota after no useful transition remains. LoopX folds the durable layer into one compact record — objective, gates, todos, scope, evidence, and quota — and drives a tick that asks whether human judgment is needed, whether a safe fallback exists, and whether the loop may continue at all. The mental model the docs offer is an agent-native Kanban for long-running work: cards carry identity, authority, evidence, and continuation, and moves are validated operators such as claim, gate, monitor, and writeback.

Multi-agent coordination is treated as a peer problem rather than a hierarchy. Registered agents hold claims, leases, task boundaries, and capabilities, and typed continuation decides who acts next — there is no durable leader identity. The host matrix is broad: Codex App and Codex CLI, Claude Code via an opt-in adapter that gates the native /loop, OpenCode, Pi, and Cursor or any shell runner. The core tick is deliberately small enough to embed by hand — quota should-run, todo claim, todo update, refresh-state, quota spend-slot.

Installation is genuinely lightweight. Python 3.11+ with no runtime dependencies outside the standard library, installable via a curl script without cloning, after which loopx connect and loopx doctor wire up a project. State lives in .loopx/ and is meant to be gitignored rather than committed.

The evidence presentation deserves credit for restraint. The headline showcases — a 200+ hour public OpenViking contribution arc and a redacted owner-run AutoML experiment of similar span — are labeled as elapsed wall-clock lifetime, not continuous model execution, and the project states plainly that these are demo results and user reports rather than reproduced production outcomes. Independent-user cases (a four-day unattended run, seven merged PRs) are marked as attribution that remains unverified.

The caveats follow from the same honesty. LoopX states it is not an autonomous production controller — dangerous permissions, publishing, production writes, and final ownership stay with a human — and the status badge still reads loop agents early. Documentation is bilingual but Chinese-first in places, with the user manual hosted on Feishu, which is friction for some readers. MIT licensed.