LoopX
An open, local-first control plane that keeps long-horizon coding and research agents working across sessions, tracking goals, decisions, and evidence.
LoopX runs alongside a coding agent runtime, such as Codex, Claude Code, or DeepSeek Harness, keeping track of what to do next, what has been accepted, and when to stop and ask the user, rather than providing the model or tools itself. Its Personal Agent Workspace keeps goals, conversations, tasks, files, schedules, and recovery state durable across restarts and across different harnesses, so reopening a project shows the previous turn's state and evidence and lets work continue from the next permitted action.
A user can see what needs their attention, configure per-goal capability overrides before applying them, steer a live turn or queue a message from a connected chat integration, inspect deliverable files and their supporting evidence, and review protected changes through an explicit preview and confirmation step, with LoopX's own state, not the browser, treated as authoritative. On a 46-task benchmark, the project reports its scheduler improving mean reward over both plain and goal-aware Codex baselines.
It is written in Python and distributed under the Apache-2.0 license.