Skip to main content
Your coding agent can use LithosBox to run commands, install dependencies, inspect results, and save checkpoints in an isolated environment. Install the LithosBox skill so the agent has the setup instructions and API guide available in its session.

​
Install the skill

Run the installer in your terminal:
Select Claude Code, Codex, or Cursor when prompted. Enter an API key from Console → API Keys and choose whether to save it. The optional sandbox check verifies that your account can create, use, and delete a sandbox. Restart your agent or reload its skills after installation. The installed runtime.md tells it which local command to use for the SDK and saved credentials. See Installation to choose agents explicitly or install for a single project.

​
Ask your agent to use a sandbox

Describe the task, the inputs it should use, and what you want to keep. For example:
For work you want to continue later:
The agent runs its SDK script on the computer where you installed LithosBox. Commands sent through the SDK run inside the sandbox. Files from your project need to be uploaded before the sandbox can use them.

​
Continue work in another session

Keep the sandbox ID to reconnect to a working environment, or a snapshot ID to restore a saved checkpoint. Give that ID to the agent along with your next task. For parallel experiments, ask the agent to fork the sandbox after setup. Each copy starts from the same state and can be changed independently. See Save, restore, fork.

​
Check results and cleanup

Ask the agent to check command exit codes and collect output before reporting success. A background job starting successfully does not mean it has finished; its logs or completion status need to be checked separately. Delete sandboxes when the task is complete, or explicitly ask to keep one. Save a durable snapshot before deleting an environment you want to restore. Snapshots are managed separately on the console’s Snapshots page. If credentials are missing or rejected, run lithosbox configure in your terminal. You do not need to paste the key into the agent conversation.

​
Use MCP tools

If you prefer tools exposed through MCP, include the MCP dependencies when installing:
Activate the installed commands, then verify the tools are available:
For Claude Code, register the local command:
For another MCP client, set the command to the absolute bin/lithosbox path printed by the installer and use mcp as its argument. That command loads the key you chose to save. Install and configure LithosBox separately on any other computer that will run the MCP server.

​
Make the docs available to your agent

To connect the documentation search server in Claude Code:
The documentation server helps your agent look up usage; the sandbox MCP tools perform actions in your account. Agents can also read pages as Markdown by adding .md to a page URL. The documentation index and full documentation are available without connecting a tool.

​
Install skills in an existing Python environment

If you already manage an environment containing the LithosBox SDK, you can install its bundled skill directly:
This writes to .claude/skills/lithosbox/ in the current project. Use --dir ~/.agents/skills for Codex user skills or --dir ~/.cursor/skills for Cursor user skills. This command copies the skill only; follow its setup.md to configure the runtime and credentials if needed.