Agent Harbor

Sandboxed execution environment and orchestration layer for AI coding agents.

Run AI coding agents safely in isolated environments with time-travel debugging, parallel execution, and complete session recording.

Why Agent Harbor?

AI coding agents can write, execute, and modify code. Without proper isolation, they can:

• Access sensitive files (SSH keys, credentials, API tokens)
• Make unintended changes to your system
• Conflict with each other when running in parallel

Agent Harbor solves these problems with a purpose-built execution layer.

Core Capabilities

Sandboxed Execution

Every agent runs in an isolated environment:

• Read-only baseline — Host filesystem is protected
• Copy-on-write overlay — Changes are captured, never written to your actual files
• Sensitive areas shielded — Credentials directories blocked by default
• Network isolation — Egress disabled by default, opt-in when needed

Parallel Agent Execution

Launch multiple agents on the same task and keep the best result:

• Each agent works in its own filesystem branch
• No conflicts between concurrent sessions
• Compare outputs and cherry-pick the best solution

Time-Travel Debugging

Rewind to any point in an agent’s session:

• Filesystem snapshots at logical boundaries (per command, per tool use)
• Branch from any snapshot with new instructions
• Compare execution paths across branches

Session Recording

Every session is captured for replay and analysis:

• Terminal I/O recording with exact visual playback
• Filesystem state at each snapshot point
• Branch points for intervention

Filesystem Snapshot Providers

Supported Agents

Supported Terminal Environments

Quick Start

Install via Nix

nix profile install github:blocksense-network/agent-harbor

Verify installation

ah health

Start an agent

ah agent start --agent claude --prompt "Add unit tests"

Or launch the interactive TUI

ah

See Installation for detailed setup and First Task for a complete walkthrough.