Run coding agents in a locked-down local sandbox with repo-only filesystem access and controlled egress using agent-sandbox

Put Claude Code, Codex, Gemini, or other supported agent CLIs inside a persistent local sandbox instead of letting them operate directly on the host.

Prerequisites

Docker-compatible runtime, VM layer such as Colima, terminal or supported devcontainer IDE

Installation

Use the upstream install or setup path that matches your environment:

• brew install colima docker docker-compose docker-buildx

Requirements and caveats from upstream:

• Target platform: Colima + Docker Engine on Apple Silicon. Should work with any Docker-compatible runtime.

• 1. Install prerequisites

• You need a VM and Docker installed. This can be done in a variety of ways.

Basic usage or getting-started notes:

• Run AI coding agents in a locked-down local sandbox with:

• CLI (preferred) - run the agent in a terminal session using agentbox exec.

• Colima

• Source: https://github.com/mattolson/agent-sandbox

• Extracted from upstream docs: https://raw.githubusercontent.com/mattolson/agent-sandbox/HEAD/README.md

Documentation

• https://github.com/mattolson/agent-sandbox#readme

Source

• Agent Skill Exchange