Agent Harbor

If there’s a long-running, code-heavy task in your organization, you can run it with Agent Harbor.

• QA agents testing new releases on every operating system and posting bug fixes in company channels.
• Coding agents implementing a well defined spec for an entire feature or business line over days.
• Agents constantly refreshing docs with up-to-date context on your codebase.

All running in isolated workspaces locally, in the cloud or on your cluster in Agent Harbor’s copy-on-write filesystem, 100x more disk-efficient than git worktrees.

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

Four pillars of Agent Harbor: sandboxing, anti-halting, multi-OS workflows, and recording. Pick a card to dive deeper.

• Parallel orchestration without infrastructure drag

  Coordinate many concurrent agents with 10x faster startup than git worktrees and 100x lower disk usage through copy-on-write workspace branches.
• Sandboxing, secret shielding, and egress control

  Isolate filesystem writes, process trees, and network access with outbound secret tokenisation and human-approved egress and sudo brokers.
• Three-tier fleet and cross-department orchestration

  Connect workstations, servers, and auto-scaled cloud VMs over mTLS QUIC, directed by a C-suite Manager Agent across engineering, marketing, finance, and legal.
• Session recording, live attach, and time travel

  Capture every session in .ahr format with full PTY streams, filesystem snapshots at tool boundaries, and live terminal attach from any device.

How the Foundation Enables Executive-Scale Orchestration

Every capability in Agent Harbor builds directly on the layer beneath it:

1. OS-Native Sandboxing + Sub-Second Copy-on-Write Snapshots (ZFS, Btrfs, APFS, ReFS, AgentHarborFS) let you spawn dozens of parallel agents in autonomous YOLO mode in milliseconds without risking host files, credentials, port collisions, or cross-session process termination.
2. Anti-Halting Supervision + Stuck-Process Diagnosis detect hung processes, interactive prompt traps, deadlocks, infinite loops, and connection failures in real time and deliver targeted recovery feedback — making multi-hour long-running tasks possible.
3. Long-Horizon Supervisor Agents drive multi-milestone plans to completion by demanding quality, running targeted web research, rewinding and forking timelines when an approach goes off-course, and coordinating cross-platform leader/follower test runs — freeing the human from babysitting individual sessions that constantly demand input.
4. C-Suite Executive Orchestration & One-Human Departments: Because individual sessions run autonomously for hours without human hand-holding, one human can direct hundreds of concurrent agents across development, marketing, finance, and legal (“like talking to a C-suite of executives”) and own what previously required entire departments through asynchronous, text-driven collaboration (specs/*.md, milestones/*.org, issues/*.md, and .ahr recordings summarized on demand by @coordinator and @spec-index).

Three Ways to Get Started

Choose the adoption path that fits your workflow. Every path shares the same CLI (ah), recording format (.ahr), and security boundaries.

1. Install, Log In & Launch Across Your Fleet

Install the ah CLI on your workstation, authenticate with ah login, and dispatch tasks across all of your machines from agent-harbor.com , the Agent Harbor Desktop App, the Terminal UI (ah tui), or mobile and chat surfaces.

# macOS / Linux / WSL (POSIX shell)
curl -fsSL https://install.agent-harbor.com | sh
 
# Windows (PowerShell)
irm https://install.agent-harbor.com/windows | iex

Sign in to connect your machine to the fleet and synchronize your encrypted credential vault:

ah login
ah tui

See the Quickstart and Multi-Host Fleet & Cloud Executors for complete setup instructions.

2. Change Nothing — Prepend ah (ah <agent> Minimal-Impact Mode)

Keep your exact terminal workflow — including all native flags, session resumption, and model selection parameters of the third-party CLI — by prepending ah:

ah claude --resume
ah codex --model o3 --full-auto "Fix the failing unit test in auth.rs"
ah agy "Audit the API schema"
ah grok "Optimize the query planner"

When you invoke ah <agent>, Agent Harbor runs in minimal-impact mode and configures six behaviors automatically:

• Native agent TUI (--agent-ui-style native): You interact with the underlying agent’s native terminal interface and pass native flags directly or after --.
• Default local OS sandbox (--sandbox-type local): Filesystem writes, process trees, and network access are isolated by default while credential directories (~/.ssh, ~/.aws, ~/.gnupg) remain hidden.
• In-place working copy (--working-copy in-place): The agent works directly in your current repository checkout without moving paths, automatically upgrading to an isolated branch when --from-snapshot, --starting-point, or --branch is specified.
• Live anti-halting supervision: Stalled turns, interactive prompt loops, and background process hangs are detected and recovered automatically.
• Automatic remote control and session persistence (events.ahr streaming): Every run records a full terminal and event timeline (events.ahr). When a remote server is configured and you are signed in (ah login), the local daemon enrolls your workstation over QUIC and streams live sessions to the control plane so you can attach (ah session attach <id>) from any device.
• Automatic sandbox pass-through: If you disable the native sandbox or permission prompts on the command line (for example ah claude --dangerously-skip-permissions ... or ah claude ... -- --dangerously-skip-permissions / --yolo), Agent Harbor automatically disables its own sandbox (--no-sandbox) while forwarding the flag to the inner agent.

See Direct Agent Launchers & Minimal-Impact Mode for GUI launchers (--ui-mode gui), empty starts, and detached background execution (--detach).

3. Self-Host the Agent Harbor Server

The hosted service at agent-harbor.com  employs Trusted Execution Environments (TEEs) to provide hardware-level guarantees that customer data is not accessible to our team, and includes all typical enterprise features. We also provide ready-to-use cloud images for the self-hosted Agent Harbor OSS edition, which is well suited for a single organization — on AWS, GCP, Azure, Hetzner, DigitalOcean, OCI, Kubernetes, NixOS, or any Linux host:

ah server start --listen 0.0.0.0:8787
ah enroll token create --label "build-farm"

See Multi-Host Fleet, Cloud Executors & Self-Hosting for cloud image templates, NixOS modules, and PostgreSQL setup.

Documentation Map