Jonathon's AI Wiki

oh-my-claudecode (OMC) — Teams-First Multi-Agent Orchestration

9 min read

Source: Yeachan Heo Oh My Claudecode 2026 04 19

Repo: https://github.com/Yeachan-Heo/oh-my-claudecode Homepage: https://oh-my-claudecode.dev Stars: 33,085 (refresh 2026-05-08; was 30,047 at first ingest 2026-04-19, +10% / +3,038 in 19 days) Forks: 3,056 (was 2,800; +256) Latest release: v4.13.6 (May 4, 2026) — was v4.13.0 at first ingest License: MIT

Teams-first multi-agent orchestration system for Claude Code. Plugin + CLI that installs 6 orchestration modes, ~19 specialized agents, and a set of magic keywords (ralph, ulw, ralplan) on top of Claude Code. Complements Claude Code Agent Teams (Anthropic’s experimental native primitive) — OMC layers opinionated patterns and orchestration UI on top, with optional cross-model bridging to Codex and Gemini. One of the largest community multi-agent layers for Claude Code (33k stars, 3k forks, active development; 6 patch releases v4.13.1 → v4.13.6 in the 19 days after first ingest, last push 2026-05-08).

Key Takeaways

  • Teams-first design, zero-config install. The default pattern is a staged pipeline (plan → prd → exec → verify → fix), distinct from Subagents’ single-parent-fanout model. Aimed at “users who want orchestration without wiring it themselves.”
  • Six orchestration modes. Team (canonical staged pipeline), Autopilot (single-agent autonomous), Ralph (persistent verify/fix loop), Ultrawork (maximum parallelism for burst), Pipeline (sequential multi-step), omc team (tmux-based multi-worker with cross-model integration).
  • ~19 specialized agents with tier variants. Architecture, research, design, testing, data science. README references but doesn’t enumerate the full list.
  • Magic keywords as accelerators. ralph → persistent mode, ulw → Ultrawork, ralplan → plan-then-ralph chaining. Keyboard-optimized invocation for common patterns.
  • Smart model routing + HUD statusline. Haiku for simple tasks, Opus for complex. Live metrics surfaced via statusline. Cost tracking and analytics built in.
  • Cross-model orchestration. /ask codex/gemini and /ccg (tri-model synthesis) integrate Codex CLI and Gemini CLI as advisors. Bridges outside the Anthropic ecosystem in the same pattern as The Advisor Strategy — different model, different lane.
  • Native Teams opt-in. Hooks into Claude Code’s experimental Agent Teams via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. OMC is aware of and composes with the native primitive rather than replacing it.
  • OpenClaw integration. Session events forward to OpenClaw for automated agent workflows.
  • Automatic skill extraction. Runs distill into reusable Agent Skills — closes the loop between ephemeral runs and durable skill libraries.
  • Scale signal: 30k stars, 2.8k forks. One of the largest third-party Claude Code extensions. Active (last pushed same day as this ingest).
  • MIT licensed, TypeScript, npm-distributed. Published as oh-my-claude-sisyphus (naming quirk — the brand is oh-my-claudecode but the npm package is oh-my-claude-sisyphus).

Recent releases (refresh 2026-05-08)

Six patch releases shipped between first ingest (2026-04-19, v4.13.0) and refresh (2026-05-08, v4.13.6). All sourced from gh api repos/Yeachan-Heo/oh-my-claudecode/releases.

  • v4.13.6 — Reliability & macOS Hardening (2026-05-04, 14 PRs net). --madmax/--yolo now require tmux on macOS — fails fast with brew install tmux hint instead of silent direct-launch. Stop hook cleans up orphan session state (#2912 / fixes #2911). Cancel clears Ralph stop-hook artifacts (#2897). Persistent mode ignores orphan autopilot routing echoes (#2899). Launch preserves Claude auth in runtime config (#2908). The omx-teams backport (#2903) shipped and was reverted (#2910) within the window — team-runtime behavior identical to v4.13.5.
  • v4.13.5 (2026-04-28, 18 PRs / 32 commits). Team auto-merge runtime — fan-out rebase for worker branches, mailbox, leader inbox, commit-cadence, merge-orchestrator, restart recovery, gitdir safety. HUD rate-limit correctness fixes for Max / Pro / enterprise-spend edge cases; Pro/Max rate-limit preservation when subscription metadata is missing. Ralph PRD state made session-scoped. SessionStart reconciliation hardened for hard-terminated sessions. Stopped relying on hook-runner PPID for cleanup.
  • v4.13.4 (2026-04-24, 4 PRs). HUD stdin-cache scoped to session to prevent cross-session corruption (#2802). Keyword-detector blocks re-entry from pasted system-echo blocks (#2795). Native team cleanup made fail-closed (#2803).
  • v4.13.3 (2026-04-24, 15 PRs). Portable shebangs + POSIX /bin/sh fallback (#2783). .omc/skills persistence contract aligned across ignore rules + setup + docs (#2787). cleanup-orphans unref on SIGKILL escalation timer to avoid 5s CLI hang (#2774). Installer validates MCP server names before rendering Codex TOML (#2764). Centralized ultrawork protocol routing (#2761). Reduced prompt-token melting at hook ingress (#2778).
  • v4.13.2 (2026-04-22, 5 PRs). Team cross-session cancel clears owning Ralph session (#2744). Rate-limit-wait prevents stale Usage API 429s from resuming blocked panes (#2746). Codex MCP sync stops duplicating user-owned tables (#2748). HUD preserves weekly quotas when stdin rate limits present (#2751). Windows can launch npm-installed Claude CLI (#2753).
  • v4.13.1 — Cursor Support (2026-04-20, 3 PRs). cursor-agent added as 4th tmux worker type in omc-teams (executor-only, runtime-guidance test compatibility preserved) — joins Claude / Codex / Gemini in cross-model orchestration. Fixed false-positive autopilot trigger on the word “autonomous” (#2739). Self-improve artifacts scoped by topic for safer resumes (#2732).

Refresh signal. This 19-day window is dominated by reliability/correctness work — session lifecycle, cancel hooks, stale-state cleanup, HUD rate-limit fidelity, and platform hardening (macOS tmux requirement, Windows CLI launch). The one new feature is Cursor as a tmux worker type. Direction: OMC is treating “the modes work” as table stakes and pivoting to making them not break when sessions interleave or terminate weirdly. The 10% star bump in the same window (30k → 33k) is consistent with continued adoption rather than a viral spike.

How it fits the Claude Code agent stack

  • vs Agent Teams (native): OMC is higher-level opinionated patterns on top of the same primitive. Native Teams gives you multi-instance coordination; OMC gives you “team patterns for specific work shapes” (plan→exec→verify, parallelism for burst, persistent fix loops). Not mutually exclusive — OMC uses Teams under the hood when the flag is enabled.
  • vs Subagents: Subagents are isolated workers with scoped permissions, invoked by the parent. OMC’s modes orchestrate multiple subagents or teams together with a canonical pipeline — it’s the layer above.
  • vs Managed Agents: Managed Agents runs hosted, sandboxed, OAuth-enabled agents on Anthropic infrastructure. OMC runs locally via Claude Code. Different deployment surfaces; both are “how to actually run multi-agent work.”
  • vs Routines / Ultrareview: Routines and ultrareview live on Claude Code on the web (cloud-hosted). OMC is local. For a team that wants cloud-hosted automation, use Routines; for local on-demand orchestration with cross-model advisors, use OMC.
  • vs Advisor Strategy: The Advisor Strategy is Anthropic’s server-side pattern for letting smaller models consult Opus mid-request. OMC’s /ask codex/gemini and /ccg extend the same upward-consultation pattern across model families — Claude asks Codex or Gemini for a second opinion.
  • vs Agent Workflow Patterns: Anthropic’s official sequential/parallel/evaluator-optimizer pattern taxonomy — OMC is a concrete implementation that embodies multiple of these patterns in one package.
  • vs Hermes Agent: Both are multi-agent layers but Hermes is self-hosted + general-purpose; OMC is Claude-Code-specific orchestration. Different problem domains.

Implementation

  • Tool/Service: oh-my-claudecode (OMC) — TypeScript plugin + CLI for Claude Code
  • Setup — plugin route (recommended): 
    /plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode
/setup
  • Setup — npm route: 
    npm i -g oh-my-claude-sisyphus@latest
omc setup
  • Cost: Free (MIT). Usage bills to your Claude Code plan (Max/Pro) or API key. Expected high consumption given multi-agent + persistent modes; HUD statusline provides per-session tracking.
  • Integration notes:
    • Requires Claude Code CLI and Max/Pro subscription or API key.
    • Requires tmux for team / rate-limit features.
    • For cross-model orchestration, install Codex CLI and/or Gemini CLI.
    • To enable integration with native Claude Code Agent Teams, set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in ~/.claude/settings.json.
    • Forwards session events to OpenClaw gateway for cross-agent automation.
    • Notification hooks available for Telegram, Discord, Slack.
    • Automatic skill extraction composes cleanly with Agent Skills — distilled skills become reusable across projects.

Open Questions

  • Agent enumeration. README cites “19 specialized agents” but doesn’t list them. Which agents exist, what are their tier variants, and how do they map to native Agent Teams patterns?
  • Ralph mode durability. Persistent verify/fix loops can explode in cost. What are the practical stopping conditions, and how does the HUD surface runaway sessions?
  • Cross-model orchestration fidelity. /ccg (Codex + Gemini + Claude tri-model synthesis) — is the synthesis quality better than Claude-alone, and by how much? No benchmark in the README.
  • OMC vs native Teams trade-off. When should a team adopt OMC vs rely on native Agent Teams directly? OMC is more opinionated; the opinions may or may not fit a given workflow.
  • Skill-extraction quality. Automatic skill extraction is a bold claim. How good are the generated skills, and how well do they compose with Anthropic’s Skills standard (64-char name, 1024-char description, 3-tier progressive disclosure)?
  • Package naming. Why is the npm package oh-my-claude-sisyphus when the brand is oh-my-claudecode? Historical artifact or intentional split?
  • 30k stars in 3 months — stickiness unknown. Created 2026-01-09, now at 30k stars — extraordinary velocity. Unclear how many of those are working installs vs aspirational stars. Worth revisiting in 60 days.

Try It

  1. Install as a plugin in one project. Use the /plugin marketplace add … && /plugin install oh-my-claudecode && /setup flow. Keep it scoped to a single repo initially so you can diff behavior.
  2. Run /team on a multi-file feature. Pick a change that spans 3+ files and has clear “plan/implement/verify” shape. Compare the staged pipeline against your normal single-agent flow.
  3. Try /ultrawork on a burst task. Use for a task with many independent chunks (e.g., bulk refactor). Measure wall-clock vs subagents-by-hand.
  4. Test Ralph mode with a known-flaky suite. Persistent verify/fix is designed for this. Watch the HUD statusline for cost trajectory — Ralph can burn budget fast.
  5. Compose with Advisor Strategy. Run /ccg (tri-model synthesis) on a hard design question. Compare against Claude-solo with /advise. Different upward-consultation mechanisms, same ambition.
  6. Enable Agent Teams under it. Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and observe whether OMC’s Team mode uses the native primitive. Useful for understanding the composition.
  7. Wire the OpenClaw gateway. If already on OpenClaw, enable session-event forwarding so OMC runs become observable and replayable in the OpenClaw workflow.
  8. Skim the skill-extraction output after 10 sessions. See what skills OMC has auto-extracted and whether they’re worth checking into a reusable skill library.

Graph View

Backlinks