Sandboxed.sh Library Management

The Sandboxed.sh Library is a Git-backed configuration repo that stores reusable skills, agents, commands, tools, rules, MCP servers, and workspace templates. Use the library-* tools to read and update that repo.

When to Use

• Creating or updating skills, agents, commands, tools, rules, or MCPs
• Syncing library git state (status/sync/commit/push)
• Updating workspace templates or plugins in the library

When NOT to Use

• Local file operations unrelated to the library
• Running missions or managing workspace lifecycle

Tool Map (file name + export)

Tool names follow the pattern <filename>_<export>.

Skills (library-skills.ts)

• library-skills_list_skills
• library-skills_get_skill
• library-skills_save_skill
• library-skills_delete_skill

Agents (library-agents.ts)

• library-agents_list_agents
• library-agents_get_agent
• library-agents_save_agent
• library-agents_delete_agent

Commands / Tools / Rules (library-commands.ts)

• Commands: library-commands_list_commands, library-commands_get_command, library-commands_save_command, library-commands_delete_command
• Tools: library-commands_list_tools, library-commands_get_tool, library-commands_save_tool, library-commands_delete_tool
• Rules: library-commands_list_rules, library-commands_get_rule, library-commands_save_rule, library-commands_delete_rule

MCPs + Git (library-git.ts)

• MCPs: library-git_get_mcps, library-git_save_mcps
• Git: library-git_status, library-git_sync, library-git_commit, library-git_push

Procedure

1. List existing items
2. Get current content before editing
3. Save the full updated content (frontmatter + body)
4. Commit with a clear message
5. Push to sync the library remote

File Formats

Skill (skill/<name>/SKILL.md)

---
name: skill-name
description: What this skill does
---
Instructions for using this skill...

Agent (agent/<name>.md)

---
description: Agent description
mode: primary | subagent
model: provider/model-id
hidden: true | false
color: "#44BA81"
tools:
  "*": false
  "read": true
  "write": true
permission:
  edit: ask | allow | deny
  bash:
    "*": ask
rules:
  - rule-name
---
Agent system prompt...

Command (command/<name>.md)

---
description: Command description
model: provider/model-id
subtask: true | false
agent: agent-name
---
Command prompt template. Use $ARGUMENTS for user input.

Tool (tool/<name>.ts)

import { tool } from "@opencode-ai/plugin"

export const my_tool = tool({
  description: "What it does",
  args: { param: tool.schema.string().describe("Param description") },
  async execute(args) {
    return "result"
  },
})

Rule (rule/<name>.md)

---
description: Rule description
---
Rule instructions applied to agents referencing this rule.

MCPs (mcp/servers.json)

{
  "server-name": {
    "type": "local",
    "command": ["npx", "package-name"],
    "env": { "KEY": "value" },
    "enabled": true
  },
  "remote-server": {
    "type": "remote",
    "url": "https://mcp.example.com",
    "headers": { "Authorization": "Bearer token" },
    "enabled": true
  }
}

Guardrails

• Always read before updating to avoid overwrites
• Keep names lowercase (hyphens allowed) and within 1-64 chars
• Use descriptive commit messages
• Check library-git_status before pushing