toolkit

My setup for Claude Code and Codex (CLI and app), the same on each machine.

git clone git@github.com:spinsirr/toolkit.git ~/toolkit
cp ~/toolkit/.env.example ~/toolkit/.env   # add keys if mcp.json needs them
~/toolkit/install.sh

• AGENTS.md: global instructions (engineering taste, Lore memory). Tech stacks stay in each project's own AGENTS.md.
• plugin/: skills and hooks, installed as personal@toolkit.
• mcp.json, settings/: merged into each tool's config by install.sh.
• Check: python3 tests/check.py.

Agent Ensemble sandbox

In the environment:

• Skill source: spinsirr/toolkit, all skills. The skills CLI finds them in plugin/skills.

• Setup script:

  git clone -q https://github.com/spinsirr/toolkit.git ~/toolkit
  python3 ~/toolkit/scripts/apply.py ensemble

• Instructions: only the rules for Agent Ensemble. AGENTS.md has the rest, and its memory section names the service for each place: Lore on my machines, CoreSpeed Memory on Agent Ensemble.

apply.py ensemble only adds. It imports AGENTS.md from ~/.claude/CLAUDE.md, appends a copy to ~/.codex/AGENTS.md, and links each subagent that is not installed yet. The skills CLI does not install subagents, and some skills use them.

This repo is the only source. The platform can push only to corespeed-io repositories, so a session pushes a branch to the fork corespeed-io/toolkit and gives you a compare link. You open the PR here. The fork's main is not used, so nothing needs to be synced.