Claude Code Delegate

Headless Claude Code delegate skill by Artem Letyushev.

This repository packages a cross-platform agent skill that lets a controlling agent run Claude Code once, non-interactively, capture the result as JSON, and verify the work outside the delegated agent.

What It Does

• Runs claude -p through scripts/delegate_claude.py
• Captures stdout, stderr, exit code, output files, and the final response
• Supports macOS, Linux, WSL, and Windows
• Uses CLAUDE_BIN when the executable is not on PATH
• Terminates the whole child process tree on timeout
• Keeps delegated output untrusted until the controlling agent checks diffs and tests

Install

With npx skills:

npx skills add letya999/claude-code-delegate

Or copy this repository into an agent skill directory such as .agents/skills/, .claude/skills/, or ~/.codex/skills/.

Quick Start

python3 scripts/delegate_claude.py \
  --cwd "$PWD" \
  --task "Review this repository and report the highest-risk issue." \
  --timeout 45m

Windows users can run the same wrapper with py -3. See QUICKSTART.md and references/runtime-setup.md for the full cross-platform pre-flight.

Agent Skill Entry Points

• SKILL.md is the skill instruction file.
• .well-known/agent-skills/index.json is the discovery index.
• dist/claude-code-delegate.zip is the archive artifact referenced by the index.

License

MIT License. See LICENSE.