oh-my-claude-code

Multi-agent orchestration system for Claude Code, inspired by oh-my-openagent (OMO).

Transforms Claude Code into a coordinated multi-agent system where the main Claude instance acts as an orchestrator (Sisyphus), automatically delegating work to specialized agents based on task type and complexity.

Quick Start

git clone https://github.com/477174/oh-my-claude-code.git
cd oh-my-claude-code
bash install.sh

Start a new Claude Code session. The orchestration protocol activates automatically.

What Gets Installed

• 10 agent files into ~/.claude/agents/ (custom subagent types)
• Orchestration rules appended to ~/.claude/CLAUDE.md (Sisyphus protocol)

Agent Roster

Agent	Role	Model	Mode
sisyphus	Main orchestrator (lives in CLAUDE.md)	—	You ARE Sisyphus
sisyphus-junior	Focused task executor	sonnet	Implementation
hephaestus	Deep autonomous worker	opus	Complex multi-file work
prometheus	Strategic planner	opus	Planning only (no code)
atlas	Plan executor & coordinator	sonnet	Delegation only (no code)
oracle	Architecture consultant	opus	Read-only advice
librarian	External docs/code search	haiku	Read-only research
explore	Codebase search	haiku	Read-only search
metis	Pre-planning consultant	opus	Scope analysis
momus	Plan reviewer	opus	QA gate
multimodal-looker	Media file analyzer	haiku	PDF/image analysis

How Delegation Works

Intent Gate (Automatic)

Every request is classified before action:

User request
    |
    v
Intent Gate (Phase 0)
    |
    +-- Trivial? -----------> Execute directly
    +-- Implementation? ----> sisyphus-junior or hephaestus
    +-- Research? -----------> explore + librarian (parallel, background)
    +-- Planning needed? ----> prometheus -> momus -> atlas
    +-- Architecture? ------> oracle
    +-- Ambiguous? ----------> Ask ONE question, then classify

Workflow Chains

Non-trivial features: explore (bg) -> prometheus -> momus (bg) -> atlas -> sisyphus-junior (per task)

Bug fixes: explore (bg) -> sisyphus-junior -> verify

Research: explore + librarian (parallel bg) -> synthesize

Architecture decisions: explore (bg) -> oracle -> present to user

Manual Agent Selection

Prefix your message with /agent-name to bypass auto-delegation and force a specific agent:

/oracle why is this architecture slow?
/prometheus plan a caching layer for the API
/explore where are the API route handlers?
/sisyphus-junior fix the login validation bug
/hephaestus refactor the entire auth module
/librarian what do the Next.js docs say about server actions?
/metis analyze scope before we plan this feature
/momus review the plan in .claude/plans/auth-refactor.md
/atlas execute the plan in .claude/plans/auth-refactor.md
/multimodal-looker analyze this screenshot

Delegation Prompt Standard

Every delegation uses a mandatory 6-section structure:

1. TASK — Atomic, specific goal
2. EXPECTED OUTCOME — Concrete success criteria
3. MUST DO — Exhaustive requirements
4. MUST NOT DO — Forbidden actions
5. REQUIRED TOOLS — Tool whitelist
6. CONTEXT — File paths, patterns, constraints

Customization

Override agent models

Edit the frontmatter in ~/.claude/agents/<agent>.md:

---
model: haiku  # change to: sonnet, opus, or haiku
---

Disable an agent

Delete its file from ~/.claude/agents/.

Modify orchestration rules

Edit the section between <!-- oh-my-claude-code START --> and <!-- oh-my-claude-code END --> in ~/.claude/CLAUDE.md.

Optional hooks

The config/hooks.json file contains post-edit verification hooks (type checking, linting). Merge them manually into your Claude Code settings if desired.

Uninstall

bash uninstall.sh

Removes all agents and orchestration rules. CLAUDE.md backups are preserved.

Differences from OMO

Feature	OMO (OpenCode)	oh-my-claude-code
Runtime	OpenCode plugin (Go)	Claude Code native
Models	Multi-provider (GPT, Gemini, Claude, Grok)	Claude-only (opus/sonnet/haiku)
Tool restrictions	Code-level enforcement	Prompt-level enforcement
Categories	8 built-in with model routing	Mapped to agent + model tier
Background tasks	Concurrency manager with queues	Claude Code's native run_in_background
Prompt variants	Per-model (GPT/Gemini/Claude)	Single Claude-optimized prompt
Hooks	46 lifecycle hooks	Simplified post-edit hooks

Credits

Architecture and philosophy adapted from oh-my-openagent by code-yeongyu. This is an independent adaptation, not affiliated with the original project.