One config. Every project. All your AI coding agents in sync.
Claude Code + Codex + Gemini CLI — shared guidelines, skills, and workflows via symlinks. Update once, propagate everywhere.
# Clone once
git clone https://github.com/MiaoDX/claude-devkit.git ~/claude-devkit
# In any project directory
cd ~/your-project
~/claude-devkit/scripts/setup.shWorks on macOS and Linux:
 |
 |
| macOS — setup.sh | Ubuntu — update.sh |
Run setup.sh from any project directory — same command, instant symlinks:
cd ~/projects/robotics-arm && ~/claude-devkit/scripts/setup.sh
cd ~/projects/web-dashboard && ~/claude-devkit/scripts/setup.sh
cd ~/projects/ml-pipeline && ~/claude-devkit/scripts/setup.sh~/claude-devkit/scripts/update.shInstalls Claude Code, Gemini CLI, Codex, GSD, MCP servers, and skills — all in parallel.
Shared rules live in AGENT_CORE.md, with thin per-tool entry files:
CLAUDE.mdimportsAGENT_CORE.mdand keeps Claude-specific guidance smallAGENTS.mdkeeps Codex-specific guidance self-contained and focused on delegation behaviorGEMINI.mdpoints at the shared core so Gemini projects inherit the same repo constraints
The shared core covers:
- Parallel-first execution — default to delegation when workstreams are independent or verification-heavy
- Main-context protection — keep orchestration and final synthesis in the main thread
- Real tests, not stub theater — UTs should predict real behavior, not just pass
- Vis-based validation — complement numeric tests when geometry or rendering issues are easy to miss
- Repo constraints —
fetch-mcp,uv+.venv, no heavy simulations, no folder removal
Skills that teach Claude Code to orchestrate other AI tools:
| Skill | What it does |
|---|---|
| gemini | Delegate tasks to Gemini CLI — analysis, refactoring, code review |
| codex / codex-mify | Delegate tasks to Codex CLI (with optional Azure OpenAI proxy) |
| doc-keeper | Audit architecture docs for drift; auto-update stale claims |
Iterative review -> triage -> fix -> verify cycle across AI agents:
| Skill | Reviews | Reviewer |
|---|---|---|
codex-plan-ralph-refactor |
GSD phase plans | Codex |
codex-impl-ralph-refactor |
Implemented code | Codex |
agent-teams-plan-ralph-refactor |
GSD phase plans | Claude agents |
agent-teams-impl-ralph-refactor |
Implemented code | Claude agents |
Each variant runs parallel multi-angle review, auto-routes findings to the right files, persists state across sessions, and stops when issues converge to zero.
# Review plans before executing
/codex-plan-ralph-refactor 38
# Review code after implementation
/codex-impl-ralph-refactor 42 --fix-level must| Command | What it does |
|---|---|
/gsd_squash |
Squash noisy commits into clean, logical git history |
/gsd_status [N] |
Show status of last N GSD phases |
| 55+ GSD commands | Full project lifecycle — plan, execute, verify, ship |
| Script | Purpose |
|---|---|
setup.sh |
Symlink configs into any project |
update.sh |
Install/update all AI CLI tools in parallel |
convert-docs.sh |
Convert code/docs to LLM-ready markdown |
Edit AGENT_CORE.md for shared rules, then keep each tool entry file thin.
| Tool | Integration |
|---|---|
| Claude Code | Primary runtime — all skills and commands built for it |
| Codex CLI | Skills for delegation and Ralph Loop code review |
| Gemini CLI | Skill for task delegation with model selection |
| GSD (Get Shit Done) | Vendored — 55 commands, 18 specialist agents |
| MCP Servers | fetch-mcp installed via update.sh |
AI coding tools are powerful but chaotic — inconsistent configs, copy-pasted prompts, no shared patterns across projects.
claude-devkit fixes that:
- Consistency — Same guidelines across all projects, always in sync
- Speed — Skills and commands that encode what actually works
- Portability — Symlinks, not copies — update once, propagate everywhere
- Multi-agent — Claude, Codex, and Gemini working together, not in silos
Built from real usage patterns across robotics, backend, and full-stack projects. Nothing fancy, just what works.
PRs welcome — from both humans and AI agents.
Whether you're adding a new skill, improving guidelines, fixing a bug, or your AI coding agent generated a useful improvement — open a PR. Good ideas don't care who (or what) wrote them.
MIT