This skill uses the E2B sandbox CLI to let an AI agent create and manage isolated Linux environments for running code, testing packages, executing commands, and working with binary files. It is built for developers who need a safe, disposable workspace for experimental or untrusted code. The sandbox supports file uploads and downloads, browser automation for UI validation, and multi-agent setups where each agent tracks its own sandbox independently.
Follow the documented setup, then try a first task.
Installs: Repository-local Agent Sandboxes skill and bundled E2B sandbox CLI. · Claude Code, Gemini CLI, Codex CLI
Run from: Cloned repository root; use .claude/skills/agent-sandboxes/sandbox_cli/ for CLI dependency installation and CLI commands.
Clone the disler/agent-sandbox-skill repository if needed, and open its project root.
In a terminal at the project root, copy the environment template.
cp .env.sample .env
Edit the project-root .env file to set E2B_API_KEY to your own E2B API key, replacing the sample value sbx_.... Obtain the key from https://e2b.dev/dashboard/keys.
Change the terminal working directory to .claude/skills/agent-sandboxes/sandbox_cli/, then install dependencies and display CLI help.
# Install dependencies uv sync # Run the CLI uv run sbx --help
Open your agentic coding tool in the repository and start prompting it to use Agent Sandboxes.
Uses Agent Sandboxes
Use the Agent Sandboxes skill to run print(sum([1, 2, 3])) in an isolated E2B sandbox. Validate the API key first, keep the script in the sandbox, use the skill's documented timeout, and remember the sandbox ID. Report the output and sandbox ID.
Claude
Codex
Gemini CLI