English | 中文 | 日本語

invoke-opencode-acp

A skill for Claude Code CLI that saves thousands of tokens by delegating complex tasks to OpenCode subagent.

Note: This is a skill for the Claude Code CLI tool, not a standalone application. You must have Claude Code installed to use it.

Requirements: This skill requires OpenCode (via ACP protocol) and Node.js to function.

What This Skill Does

Delegates complex tasks to a separate AI agent (OpenCode) through the ACP protocol, saving your main Claude conversation from consuming excessive tokens.

Think of it as hiring a specialized contractor: you provide the objective, they do the work independently, and you get the results without all the back-and-forth that would fill your conversation history.

Benefits

Massive Token Savings

• Without this skill: Main Claude reads every file → analyzes → edits → repeats. Each iteration consumes thousands of tokens.
• With this skill: Send objective to subagent → subagent works independently → get final results. Only the summary enters your conversation.

Typical savings: 50-90% fewer tokens for complex tasks.

Important Trade-off

How it achieves savings: OpenCode subagent uses free tokens to call alternative models (e.g., GLM-4.7, Qwen) instead of your paid Claude tokens.

What this means:

• ✅ Token savings: You pay nothing for the subagent's work
• ⚠️ Capability trade-off: Alternative models have lower reasoning ability compared to Claude
• ⚠️ Quality variance: Results may be less nuanced or require review

Best use cases:

• Code refactoring and formatting (straightforward)
• Documentation updates and generation
• Test code generation
• Bug fixes with clear requirements
• Batch file operations

Consider alternatives for:

• Complex architecture decisions
• Nuanced code reviews
• Sensitive logic changes
• Tasks requiring deep domain expertise

Faster Task Completion

• Main Claude stays responsive for other tasks
• Subagent works in parallel
• No context window limitations for the subtask

Better for Complex Workflows

Perfect for:

• Multi-file refactoring: "Add type hints to all Python files"
• Batch operations: "Update all documentation files"
• Code reviews: "Review the entire codebase for security issues"
• Git workflows: "Create a pull request with these changes"

How It Works

1. You describe the task (one sentence is enough)
2. OpenCode subagent receives your objective and works independently
3. Results are saved to a file (not your conversation)
4. You get a concise summary in your main conversation

The subagent can read files, make edits, run tests, and more—just like a full developer—without consuming your main conversation's context.

Installation

Prerequisites

You need three things installed:

1. Claude Code CLI (already installed if you're using this)
2. OpenCode CLI (npm install -g opencode)
3. Node.js (comes with most systems)

Quick Install

# 1. Install OpenCode (if not already)
npm install -g opencode

# 2. Copy the skill to Claude
cp -r skills/invoke-opencode-acp ~/.claude/skills/

# 3. Verify it works (optional)
node ~/.claude/skills/invoke-opencode-acp/tests/test_acp_client.js

Verify Installation

Check that OpenCode is installed:

opencode --version

Check that the skill exists:

ls ~/.claude/skills/invoke-opencode-acp/SKILL.md

Usage

Simply tell Claude what you want done:

"Refactor all Python files in this project to use type hints"

Claude will delegate the task to OpenCode subagent and return the results.

Timeout guidelines (OpenCode is slow):

• Simple tasks (math, short answers): 180s (3min) minimum
• Medium tasks (single file, ~100 lines): 600s (10min) minimum
• Complex tasks (multi-file, refactoring): 1800s (30min) minimum

Example tasks:

• "Refactor the entire codebase"
• "Update the README.md with new instructions"
• "Add type hints to all Python files"
• "Run security audit on the codebase"

Real-World Examples

Example 1: Large Refactoring

Your request: "Add type hints to all Python files in this project"

Without this skill:

• Claude reads each file (10,000 tokens)
• Claude edits each file (15,000 tokens)
• Claude verifies changes (8,000 tokens)
• Total: ~33,000 tokens consumed

With this skill:

• You send the objective (50 tokens)
• Subagent works independently (0 tokens in your conversation)
• You get a summary (200 tokens)
• Total: ~250 tokens consumed

Savings: 99% token reduction!

Example 2: Code Review

Your request: "Review the entire codebase for security vulnerabilities"

Without this skill:

• Claude reads every source file (25,000 tokens)
• Claude analyzes and documents (12,000 tokens)
• Total: ~37,000 tokens

With this skill:

• Objective + summary (~300 tokens)
• Total: ~300 tokens

Savings: 99% token reduction!

Example 3: Batch Documentation Updates

Your request: "Update all .md files to include the new license header"

Without this skill:

• Claude reads 20 markdown files (15,000 tokens)
• Claude edits each file (20,000 tokens)
• Total: ~35,000 tokens

With this skill:

• Objective + summary (~250 tokens)
• Total: ~250 tokens

Savings: 99% token reduction!

When to Use

✅ Use this skill for:

• Tasks affecting 2+ files
• Refactoring or restructuring
• Batch operations (find and replace across many files)
• Code reviews across the entire codebase
• Multi-step tasks requiring research and analysis
• Git workflows (commit, push, pull request)

❌ Don't use for:

• Single file quick edits
• Simple text replacements
• Tasks requiring integration with your main conversation

How This Is Different

vs. Normal Claude Tasks

Aspect	Normal Claude	With invoke-opencode-acp
Context usage	Full file contents	Just objective
Token consumption	High (thousands)	Low (hundreds)
Your conversation	Shows all steps	Shows only results
Speed	Sequential	Parallel
Best for	Simple tasks	Complex tasks

vs. Other Solutions

• Claude Projects: Still uses main context window (limited tokens)
• Manual delegation: Requires copy-pasting, no automation
• invoke-opencode-acp: Automated delegation with minimal context usage

Troubleshooting

"opencode command not found"

npm install -g opencode

"Skill not enabled"

• Check the skill is in ~/.claude/skills/invoke-opencode-acp/
• Restart Claude Code
• Enable in Claude Code settings

Task timeout

• Increase the -t parameter value
• See timeout guidelines in Usage section
• OpenCode can be slow, especially for complex tasks

Technical Details (For the Curious)

This skill uses the ACP (Agent Control Protocol) to communicate with OpenCode:

1. Launches opencode acp via Node.js child_process
2. Creates a dedicated session via JSON-RPC
3. Sends your task via session/prompt
4. Streams results via session/update events
5. Filters out <thinking> tags
6. Returns only the final output

Implementation: skills/invoke-opencode-acp/acp_client.js (192 lines)

License

MIT License - See LICENSE for details.

---

中文版本：README.zh.md 日本語版本：README.ja.md