Agent Harbor
Sandboxed execution environment and orchestration layer for AI coding agents.
Agent Harbor lets you run AI coding agents (Claude, Gemini, Codex, Copilot, Cursor) safely in isolated environments. Without it, agents can access sensitive files, make unintended changes, or conflict when run in parallel. Agent Harbor gives you a purpose-built execution layer: copy-on-write filesystem, time-travel debugging, and full session recording so you can let agents run for hours on real tasks.
You get a TUI for interactive task and session management, a CLI for automation and CI, and support for AgentFS, ZFS, and Btrfs snapshot providers. For supported agents, terminals, quick start, and filesystem options, see Under the hood.
TUI at a glance
The interactive TUI is the main way to manage tasks and sessions. Launch it with ah and create tasks, watch runs, and replay sessions from one place.
Core features
Five pillars of Agent Harbor: sandboxing, anti-halting, multi-OS workflows, time travel, and recording. Pick a card to dive deeper.
Anti-halting keeps parallel runs moving
When one branch fails, supervisors detect it, rollback automatically, and keep the broader workflow progressing. Self-healing agents with targeted web research recover without manual babysitting.