OpenCode Intern Skill

A Claude Code skill that delegates tasks to an OpenCode "intern" agent via ACP (Agent Communication Protocol). This enables Claude Code to offload tasks to a secondary AI agent running in OpenCode using any model you like.

OpenCode can read AGENTS.md in the same manner Claude Code does, so you can seamlessly share project context between both systems. You could even point the OpenCode Rules to the CLAUDE.md file

An added benefit of this setup is OpenCode automatically uses skills in the .claude/skills directory so there's no need for duplicating files, and both agents can share the same working space and methodologies.

To avoid nested callbacks that Claude looses sight of, make sure to deny permission to the OpenCode-intern skill in the OpenCode config.

For more portability and less token use, you could transpose a basic prompt in to the OpenCode.json file directly rather than point to a markdown file.

Prerequisites

• Claude Code installed
• OpenCode installed and available in PATH
• Python 3.8+
• The OpenCode.json is currently configured to use the free OpenCode Zen Minimax M2.1 model through Zen, but obviously you can configure it with any you like.

Installation

1. Copy the OpenCode configuration

Copy the following to your project root:

OpenCode.json          -> your-project/OpenCode.json
.OpenCode/             -> your-project/.OpenCode/

2. Copy the Claude Code skill

Copy the skill folder to your Claude Code skills directory:

.claude/skills/OpenCode-intern/  -> your-project/.claude/skills/OpenCode-intern/

Or to your global skills directory:

• macOS/Linux: ~/.claude/skills/OpenCode-intern/
• Windows: %USERPROFILE%\.claude\skills\OpenCode-intern\

The default configuration uses the free OpenCode Zen model, so no API key is required to get started.

Usage

In Claude Code, use phrases like:

• "Use the intern to refactor this function"
• "Ask the intern to write tests for this module"
• "Delegate this task to the intern"

Claude Code will invoke the skill, which communicates with OpenCode via ACP to execute the task.

File Structure

OpenCode-intern-skill/
├── README.md                           # This file
├── OpenCode.json                       # OpenCode configuration
├── .OpenCode/
│   └── prompts/
│       └── intern.md                   # Intern agent system prompt
└── .claude/
    └── skills/
        └── OpenCode-intern/
            ├── skill.md                # Skill definition
            └── scripts/
                └── acp_client.py       # ACP protocol client

Configuration

Intern Agent Permissions

The intern agent has a minimal permission configuration:

"permission": {
    "skill": {
        "OpenCode-intern": "deny"
    }
}

The OpenCode-intern: deny entry prevents recursive self-calls. You can extend this to allow or deny other skills and bash commands as needed.

Intern System Prompt

The intern agent's behaviour is defined in .OpenCode/prompts/intern.md. Key characteristics:

• Precision over initiative: Executes exactly what is specified
• No autonomous decisions: Reports blockers instead of making assumptions
• Scope boundaries: Works strictly within defined parameters
• Silent efficiency: Minimal commentary unless requested

ACP Client

The acp_client.py script handles communication with OpenCode:

python acp_client.py <project_path> "<prompt>" --json

Parameters

Parameter	Description
project_path	Absolute path to the project directory
prompt	The task to delegate to the intern
--agent	Agent name (default: "intern")
--timeout	Timeout in seconds (default: 300)
--json	Output results as JSON

Response Format

{
    "success": true,
    "response": "Agent's response text",
    "tools": [
        {"name": "bash", "status": "completed"},
        {"name": "read", "status": "completed"}
    ],
    "errors": []
}

Modifying the Intern's Behaviour

Edit .OpenCode/prompts/intern.md to adjust:

• Execution philosophy
• Processing protocol
• Ambiguity handling

Inline Prompts

For smaller token usage, you can embed the prompt directly in OpenCode.json instead of referencing a file:

{
    "agent": {
        "intern": {
            "prompt": "You are the Intern. Execute tasks precisely as specified. Report blockers instead of making assumptions."
        }
    }
}

Troubleshooting

Error	Solution
"OpenCode not found"	Ensure OpenCode is installed and in PATH
"Timeout"	Increase --timeout or break task into smaller pieces
"Session error"	Check OpenCode logs
"Init failed"	Verify OpenCode version supports ACP protocol