oh-my-openagent (OmO) — Full Review:
"The Multi-Model Harness
That Escaped Claude's Prison"
An OpenCode plugin that transforms a single AI coding assistant into a virtual dev team of 11 specialized agents working in parallel. 37.6K GitHub stars, a history of being blocked by Anthropic, and an open-source rebellion built by a Korean developer. Full breakdown — from install to real-world use.
What Is oh-my-openagent (OmO)?
OmO is a multi-model agent orchestration harness that sits on top of OpenCode. It transforms a single AI coding assistant into a virtual dev team of 11 specialized agents working in parallel. Claude, GPT, Gemini, Kimi, GLM, Minimax — instead of being locked to one model, OmO automatically routes each task to the most suitable model.
The project goes by "omo" or "Sisyphus". It was previously called oh-my-opencode and recently rebranded. The creator is Korean developer @code-yeongyu. Remarkably, 99% of this project was built using OpenCode itself — AI building AI tooling.
Claude Code is a nice prison. But it's still a prison. We don't do lock-in here. We ride every model.
— oh-my-openagent README, 2026The core philosophy is "model independence." No single provider will dominate forever — models get cheaper and smarter every month. OmO is designed for that open market. Not for walled gardens.
Origin Story: The Project Anthropic Blocked
OmO has a dramatic backstory. Anthropic blocked OpenCode's API access specifically because of this project. This has been officially confirmed. The team's response was not to comply — it was to fight back.
The result: Hephaestus — an agent named "The Legitimate Craftsman." Built on GPT-5.3 Codex, it operates entirely independently of Anthropic. The irony in the name is deliberate. In Greek mythology, Hephaestus was banished from the halls of the gods — yet forged the finest things in existence.
"Anthropic blocked OpenCode because of us. That's why Hephaestus is called 'The Legitimate Craftsman.' The irony is intentional."
The development process is equally unconventional. The maintainer builds and maintains every feature, bug fix, and issue triage live on Discord — using Jobdori, a heavily customized OpenClaw-based AI assistant. This is a project publicly experimenting with AI building AI tooling in real time.
The 11 Specialist Agents: Meet Your Virtual Dev Team
OmO's core idea is role separation. Instead of one AI handling everything, each specialized agent handles its domain with an optimized model. Sisyphus conducts the whole orchestra; the others execute.
Plans, delegates, and executes complex tasks via parallel background agents. Todo-driven workflow with a 32k extended thinking budget. Activated at full power with the ultrawork keyword.
Inspired by AmpCode's deep mode. Explores codebase patterns, researches context, and completes tasks end-to-end without premature stopping. Born from Anthropic's block — "The Legitimate Craftsman."
GPT-5.3 Codex (dedicated, no fallback)Enter with Tab. Conducts an interview-style planning session before implementation. "Fire brought by the gods" — sets direction before action. Launches full orchestration via /start-work.
Architecture decisions, code review, and debugging. Read-only consultation with deep logical reasoning and analysis. Inspired by AmpCode.
GPT-5.4 high / Gemini-3.1-pro fallbackDedicated to official docs, GitHub code search, and web research. Uses Context7, Grep.app, and Exa MCP tools. Called by Sisyphus before implementation to gather relevant context.
Compatible with local Ollama modelsDedicated to understanding filesystem and codebase structure before implementation. Runs on lightweight local models to minimize cost.
Ollama qwen2.5-coder:7b compatibleSpecialized in UI/UX, CSS, and visual components. Has access to browser automation tools via Playwright MCP.
Gemini (for creativity)Critical review of implemented code and patch validation. Assigned the highest-reasoning model available. Acts as the quality gate before release.
GPT-5.4 xhigh (top reasoning)※ The remaining 3 agents (GPTPhus, Sisyphus-Junior, Big-Pickle) serve special-purpose and cost-optimization roles.
How to Use: From Install to ultrawork
The official recommendation is not to read the installation docs yourself — paste the URL into an agent and let it install OmO for you. That philosophy tells you everything about the project. But here's the step-by-step for those who want to understand what's happening.
oh-my-opencode even though the repo is now oh-my-openagent.omo from your project root. OmO automatically reads AGENTS.md and CLAUDE.md from your project to inject context — run it from the right directory.ultrawork or ulw inside the TUI. Sisyphus spins up the full team in parallel — research, implementation, and verification happening simultaneously.Use Cases: When OmO Actually Shines
"Oh My Opencode delivers on its promise for the right workload. For the wrong one, it adds cost and complexity without improving results. Knowing the difference is the whole game."
Where OmO wins (parallel work, multi-file): features spanning many files, integrating unfamiliar libraries, large-scale refactors, bugs tangled across multiple modules, and tasks requiring research + design + implementation + verification in sequence. In real testing, an IndexNow API integration task — where Librarian gathered official docs while Hephaestus implemented — produced substantially better results than a bare model run.
Where vanilla OpenCode is better (single context): single-file edits, quick questions, document analysis, and pure reasoning tasks that fit in one context window. One real-world test showed identical results to vanilla OpenCode on a document migration task — but OmO burned roughly 3× the tokens to get there.
- Feature spans 10+ files
- Integrating an unfamiliar library or framework
- Full codebase refactor
- Complex bugs involving multiple modules
- Research → design → implement → verify pipeline
- Cost optimization via local + cloud model mix
- Single-file edits or quick questions
- Reasoning fits in one context window
- Document analysis or summarization
- Running a powerful model like Opus (less delta)
- Budget is tight
- First time using an AI coding tool
How OmO Compares to the Competition
OmO's position is unique. It's not a replacement for Claude Code or Cursor — it's an orchestration layer on top of OpenCode. The correct comparison set is tools that operate at the same abstraction level.
| Tool | Type | Multi-model | Parallel exec | LSP / AST | Claude compat | License |
|---|---|---|---|---|---|---|
| oh-my-openagent | OpenCode harness | ✅ Auto-routing | ✅ Background | ✅ IDE-grade | ✅ Full | Open |
| Claude Code (vanilla) | Single agent | ❌ Claude only | ❌ Sequential | ❌ | ✅ Native | Commercial |
| AmpCode | Claude Code fork | ⚠ Limited | ⚠ Deep mode | ⚠ Partial | ✅ Compatible | Open |
| Cursor | IDE-integrated agent | ✅ Multiple | ❌ Single | ✅ Built-in IDE | ⚠ Partial | Commercial |
| OpenClaw | Autonomous agent | ⚠ Config needed | ❌ | ❌ | ✅ Compatible | Open |
| Paperclip | Agent org platform | ✅ Adapter-based | ✅ Task-level | ❌ | ✅ Adapter | MIT |
| Devin (commercial) | Fully autonomous dev | ✅ Internal | ✅ | ✅ | ❌ | Commercial |
OmO vs Paperclip: often confused, but different layers entirely. OmO is an agent harness specialized for coding task execution. Paperclip is a control plane for orchestrating business operations. They're complementary — you could connect OmO as one of Paperclip's agents, letting OmO handle all coding while Paperclip ties its work to company-level goals.
What's Next: The Ambitions of Sisyphus Labs
The team has already announced their next step. Under the name Sisyphus Labs, they're fully productizing OmO to "define the future of frontier agents." A waitlist is open now.
The future isn't picking one winner — it's orchestrating them all. Models get cheaper every month. Smarter every month. We're building for that open market.
— oh-my-openagent README, 2026Particularly worth watching: the team claims that Kimi K2.5 + GPT-5.3 Codex combinations are already outperforming vanilla Claude Code on certain benchmarks. As the industry shifts from single-model to multi-model orchestration, OmO is one of the earliest projects actively stress-testing that transition in production.
- Real billing incident on record: A confirmed bug in March 2026 (issue #2571) charged one user $438 in a single afternoon. Three compounding failures: no sub-agent step limit, loop detection failure, and delayed cost alerts. Fix is in progress but not shipped yet.
- High startup token overhead: A "Hello world" message consumes 15,000–25,000 tokens just to set up context. Using ultrawork on simple tasks is wasteful by definition. Deferred loading is in development but not released.
- Weaker models can backfire: Strong models like Claude Opus push through heavy orchestration layers and still produce good output. Weaker models are more sensitive — the orchestration overhead can actively hurt performance rather than help it.
- OpenCode dependency risk: OmO is a plugin, not a standalone tool. Changes to OpenCode can break OmO features. After OpenCode 1.1.37+, several features were automatically disabled as a result of upstream changes.
- Docs lag behind code: Fast development means documentation occasionally falls behind reality. The package name (oh-my-opencode) vs repo name (oh-my-openagent) mismatch is a common source of confusion for first-timers.
- Single-maintainer risk: Core development is heavily dependent on @code-yeongyu. The Sisyphus Labs incorporation partially mitigates this, but it's still early days for that structure.
- Multi-model automatic routing is among the most complete implementations in open-source coding agents today
- Full Claude Code compatibility — bring your existing skills, MCPs, and hooks with zero reconfiguration
- Hashline editing accuracy improvement is benchmark-verified (6.7% → 68.3%)
- Official Korean README and Discord community — unusually developer-friendly for a project of this scope
- Building a GPT-native alternative after Anthropic's block signals a genuine long-term commitment to independence
- Local Ollama model support enables hybrid cloud/local cost optimization — you're not 100% cloud-dependent
Final Verdict: Powerful When Used Right
OmO isn't overhyped — and it isn't a gimmick. On the right workload, used correctly, it's genuinely more capable than vanilla Claude Code. The prerequisite is understanding when that's true and when it isn't. Using ultrawork for everything is as inefficient as using a sledgehammer to hang a picture frame.
From a 22B Labs perspective, what's most interesting about this project isn't just the technical capability. It's the existence of it at all — AI building AI tooling, pushing back against Anthropic's block, and actively experimenting with the multi-model future. That direction is exactly what The 4th Path is paying attention to.
⭐ 8.4 / 10 · Recommended for: multi-file developers, multi-model explorers, cost-conscious builders
- code-yeongyu/oh-my-openagent — README.md, docs/guide/overview.md, docs/reference/features.md · github.com
- Rost Glukhov — Oh My Opencode Review: Honest Results, Billing Risks (Mar 2026) · dev.to
- Rost Glukhov — Oh My Opencode Specialised Agents Deep Dive (Mar 2026) · glukhov.org
- LobeHub Skills Marketplace — oh-my-openagent-omo skill (2026) · lobehub.com
- Termo.ai — Oh-my-openagent AI Skill (Mar 2026) · termo.ai
- oh-my-openagent/releases — Release notes and changelog · github.com
- Sisyphus Labs — Official waitlist page · sisyphuslabs.ai