LoopX

LoopX is a state kernel and local-first control plane for long-running agent work. It sits on top of an agent harness rather than replacing it: the harness executes bounded turns, while LoopX holds the durable state that spans them — the objective, gates, todos, scope, evidence, and quota. Work started days earlier resumes where the last turn stopped instead of depending on chat memory and a timer.

The loop is explicit. From the current state, LoopX decides whether human judgment is needed — in which case it asks a concrete question and waits — or whether a safe fallback exists, in which case the harness runs one bounded slice; the slice writes evidence, a handoff, and the next todo, and a quota decides whether another tick should follow. The project describes the model as an agent-native Kanban board for long-running work: cards carry identity, authority, evidence, and continuation, moves are validated operators such as claim, gate, monitor, and writeback, and the board is a projection of the state rather than the source of truth. Registered agents act as peers, with claims, leases, task boundaries, and capabilities deciding who acts next, so no durable leader identity is required.

A local-first workspace collects goals, attention, conversations, tasks, files, schedules, and recovery in one place, launched with loopx dashboard as a browser or progressive web app; an experimental Tauri desktop shell reuses the same loopback services without becoming a second state authority. Protected changes are reviewed through a typed preview, explicit confirmation, and receipts.

Host support is broad and each host keeps its own loop driver: Codex App and Codex CLI, Claude Code through an opt-in adapter, KunlunCode, OpenCode, Pi, ZCode, the Antigravity CLI, the DeepSeek Harness, and Cursor, a shell, or a custom runner. The core tick a custom runner has to implement is small — quota should-run, todo claim, todo update, refresh-state, and quota spend-slot. Installation needs Python 3.11 or later, and the package has no runtime dependencies outside the standard library.

The architecture separates a Kernel that owns durable goal, todo, gate, evidence, quota, and scheduling truth, Capabilities that define provider-neutral contracts producing one bounded verifiable outcome, Providers that call external or local systems and return bounded observations, and Extensions that package a provider with an install, readiness, enable, upgrade, disable, and rollback lifecycle. Dangerous permissions, publishing, production writes, and final ownership stay with the person running it.

Features

• Durable goal state: objective, scope, authority, gates, todos, evidence, and quota persist across turns, restarts, and harnesses
• Harness-agnostic: Codex App and CLI, Claude Code, KunlunCode, OpenCode, Pi, ZCode, Antigravity CLI, DeepSeek Harness, Cursor, and custom runners
• User gates: a concrete question addressed to a person, rather than an unexplained wait for the owner
• Quota-gated ticks: quota should-run decides whether a registered agent acts now, ending spend when no useful transition remains
• Peer agents: claims, leases, task boundaries, and typed continuation determine ownership without a fixed leader
• Personal workspace: loopx dashboard shows what needs you, what is running, what is watched, and what is scheduled or stopped
• Protected actions: typed preview, explicit confirmation, and receipts before a protected change lands
• Small integration surface: five commands (quota should-run, todo claim, todo update, refresh-state, quota spend-slot) embed the loop in any runner
• Layered extension model: Kernel, Capability, Provider, and Extension boundaries with an explicit extension lifecycle
• Standard-library only: Python 3.11 or later with no runtime dependencies beyond the standard library, installable from PyPI