oh-my-claude
____ _ _ __ ____ __ _____ _ _
/ __ \| | | | | \/ \ \ / / / ____| | | |
| | | | |__| |_____| \ / |\ \_/ /_____| | | | __ _ _ _ __| | ___
| | | | __ |_____| |\/| | \ /______| | | |/ _` | | | |/ _` |/ _ \
| |__| | | | | | | | | | | | |____| | (_| | |_| | (_| | __/
\____/|_| |_| |_| |_| |_| \_____|_|\__,_|\__,_|\__,_|\___|
[!CAUTION] EXPERIMENTAL PROJECT - This plugin runs autonomous AI loops that can consume significant tokens. Mind your token usage!
/ultraworkand/deepworkwill iterate until completion. Use--max-iterationsto set limits. You have been warned.
Claude Code plugin for AI-powered iterative development loops.
Inspired by oh-my-opencode and Ralph Wiggum.
Installation
/plugin marketplace add 2lab-ai/oh-my-claude
/plugin install oh-my-claude@oh-my-claude
/plugin install powertoy@oh-my-claude
Required: Run
/oh-my-claude:setupafter installation.
Authentication
oh-my-claude uses your Gemini CLI and Codex CLI subscriptions (not API keys).
# 1. Run setup to check dependencies
/oh-my-claude:setup
# 2. Login to each service
gemini auth login # Opens browser for Google auth
codex auth # Opens browser for OpenAI auth
# 3. Verify everything works
/oh-my-claude:setup
Once authenticated, agents use your existing subscriptions:
- Explore agent → Your Gemini subscription
- Oracle agent → Your OpenAI/Codex subscription
- Librarian agent → Your Claude subscription (already authenticated via Claude Code)
Main Features
/ultrawork - Multi-Agent Work Loop
Ralph loop with 3 specialized AI agents for autonomous development.
[!IMPORTANT] Requires YOLO mode:
claude --dangerously-skip-permissions
/ultrawork "Build REST API for users"
Agents:
- Oracle (Codex GPT-5.2): Architecture decisions, failure analysis (blocking)
- Explore (Gemini): Internal codebase search (parallel)
- Librarian (Opus 4.5): External docs, GitHub source analysis (parallel)
Loop exits when task is genuinely complete.
/deepwork - Reviewed Work Loop
Same as ultrawork, but requires 9.5+ score from all three reviewers (GPT-5.2, Gemini-3, Opus-4.5) before completion.
/deepwork "Critical security fix" --max-iterations 50
/save & /load - Cross-Tool Context Management
Save work context in one tool and resume in another.
# In Claude Code: save your work
/save
# In Gemini CLI or Codex: resume the work
/load
Cross-tool workflow example:
- Start work in Claude Code,
/savewhen done - Open Gemini CLI or Codex,
/loadto continue - Work across tools seamlessly - they share
./docs/tasks/save/
To enable cross-tool commands:
./install-cross-session-commands.sh
This installs /save, /load, /list-saves, /check to:
~/.gemini/commands/(Gemini CLI)~/.codex/prompts/(Codex)
Commands Reference
oh-my-claude
| Command | Description |
|---|---|
/ultrawork | Multi-agent autonomous work loop |
/deepwork | Work loop with triple AI review gate (GPT-5.2 + Gemini-3 + Opus-4.5, all ≥9.5) |
/save | Save work context |
/load <id> | Load saved context |
/list-saves | List all saved contexts |
/check [all|id] | Verify archived saves completion |
/cancel-work | Cancel active loop |
/setup | Verify dependencies |
powertoy
| Hook | Description |
|---|---|
| auto-title.sh | Auto-generate session titles (Claude Haiku) |
| play-sound.sh | Play sound on session end (macOS) |
MCP Servers
Bundled MCP servers:
| Server | Package |
|---|---|
| gemini | @2lab.ai/gemini-mcp-server |
| claude | @2lab.ai/claude-mcp-server |
| codex | codex mcp-server |
Credits
- oh-my-opencode: code-yeongyu
- Ralpy Wiggum plugin: https://github.com/anthropics/claude-code/tree/main/plugins/ralph-wiggum
- Ralph Wiggum technique: Geoffrey Huntley
FAQ
Q: Will this burn my tokens?
A: Yes. That's the feature, not a bug.
/\_/\
( o.o ) "I'm helping!"
> ^ < - Your AI, iteration 47
Changelog
| Version | Changes |
|---|---|
| 1.0.22 | Context window optimization - total plugin under 4k tokens |
TODO
- LSP support
License
MIT