Acp Agent Runner

Use this skill whenever the user wants a coding task carried out by an AI coding agent or model that is NOT Claude — for example OpenCode, GLM, Qwen, DeepSeek, Kimi, MiniMax, "another coding agent", "another vendor's model", or "whatever external agent is set up". Typical intents: hand or throw a task to that other agent and bring back its result; ask it to review changes, fix failing tests, or refactor; route work "through opencode"; or run the same job on two outside models to compare their outputs. This skill runs that external agent for you (over the Agent Client Protocol). Two main workspace modes: **inplace** for ordinary work handoff, **isolated** (worktree) for verifying that a skill/procedure can be completed by an external model (e.g. OpenCode Go). It applies in any language (English, 日本語, など) and to vague or indirect phrasing. Do NOT use it for Claude's own subagents (use the Agent tool instead), for plain Anthropic/LLM API scripts, or for running Claude itself — the defining signal is that the work goes to a different, non-Claude agent.

mkdir -p ~/.claude/skills/acp-agent-runner
curl -L https://claudskills.com/skills/acp-agent-runner/SKILL.md \
  -o ~/.claude/skills/acp-agent-runner/SKILL.md

@misc{acp-agent-runner-2026,
  author    = {zonuexe},
  title     = {Acp Agent Runner [Claude Code skill]},
  year      = {2026},
  publisher = {ClaudSkills},
  url       = {https://claudskills.com/skills/acp-agent-runner/}
}