The sandbox skill

For AI agents: the documentation index is at llms.txt (full corpus: llms-full.txt). A markdown source for this page is /sandboxes/skill.md.

A skill is a set of instructions an agent loads when a task calls for it. arete-sandbox teaches an agent to act from a sandbox:

• find its sandbox with sandbox_list, and choose one with sandbox_select when several are granted;
• read sandbox_status and sandbox_rules before acting;
• preview every action with sandbox_preview, then run it with sandbox_invoke;
• pass amounts as decimal strings in the smallest unit;
• follow a refusal’s hint, and never route around a rule;
• after a timeout, look the action up with sandbox_action before retrying;
• answer “what did you do” from sandbox_history; and
• never ask for a private key or the owner’s API key.

The agent loads it when the sandbox_* tools are listed, or you ask it to trade, swap, withdraw, read through a sandbox or explain a refusal.

Install

a4 init installs it, with the MCP configuration, for the agent running the command and any agent configured in the project:

a4 init -y

To install the skill on its own:

npx skills add AreteA4/skills --skill arete-sandbox
npx skills add AreteA4/skills --skill arete-sandbox --agent claude-code

Skills follow your account

a4 init installs only the skills your account’s command groups cover, the same groups that decide which a4 commands are listed. An agent account with sandbox access gets:

Skill	Group	Use
arete-sandbox	sandboxes	Acting from a sandbox
arete	registry	Discovery, exact descriptors and project dependencies
arete-programs	registry	Program accounts, PDAs and transactions

Accounts in the private beta also get the skills for streams, stacks, protocols and dev environments; see Agent skills. When the API can’t be reached, or a4 init runs without an account key, it installs every skill.

Run a4 init -y again after your account gains access to more, and a4 doctor reports any skill your account should have that is missing.