code-yeongyu/oh-my-openagent is a TypeScript OpenCode plugin (not a standalone agent harness) with 67,728 stars on GitHub (real-time, per GitHub), 5,538 forks, created 2025-12-03, and last pushed today (2026-08-12). It is not archived; its homepage is omo.dev. The npm package is oh-my-opencode, currently v5.0.0-beta.6. The package.json description reads: "The Best AI Agent Harness - Batteries-Included OpenCode Plugin with Multi-Model Orchestration, Parallel Background Agents, and Crafted LSP/AST Tools." The differentiator: unlike Claude Code (locked to Anthropic) or Cursor (closed-source IDE copilot), it acts as a model orchestration layer -- pulling multiple models (Kimi K3, GPT-5.6 Sol, and others) into a single Team Mode frame to work in concert. The README fires its sharpest hook right away: "Anthropic blocked OpenCode because of us. Yes, this is true." But the catch is right there too: the license is not MIT or Apache, it is SUL-1.0 (Sustainable Use License), a non-OSI license with competitive-use restrictions. This piece unpacks that and several other pitfalls.
This article is based on code-yeongyu/oh-my-openagent's GitHub README and LICENSE.md, accurate as of 2026-08-12; star counts move in real time. It covers different ground from this site's crush resource, pi resource, and openclaw resource, and complements but stands apart from the AI coding plan comparison.
1. Positioning: a batteries-included OpenCode plugin, not another standalone agent
Get one thing straight first: oh-my-openagent is not a standalone coding agent built from scratch -- it is a plugin for OpenCode. OpenCode itself is an open-source AI coding framework (terminal-based), and oh-my-openagent layers "batteries" on top of it: multi-model orchestration, parallel background agents, and crafted LSP/AST tools, ready to go on install. This differs from crush (a standalone terminal agent) and Claude Code (Anthropic's official closed-source agent): it does not reinvent the wheel, it uses OpenCode as the base and stacks capabilities on top.
Why does this positioning matter? Because OpenCode users already have the code infrastructure; what they lack is "orchestrating multiple models" and "running agents in parallel in the background." oh-my-openagent fills exactly that gap. There are five bin commands: oh-my-opencode, oh-my-openagent, omo-agent-toolkit, lazycodex, and lazycodex-ai -- the first three are the OmO mainline, the last two are LazyCodex (OmO for Codex, installed via npx lazycodex-ai install, site at lazycodex.ai).
The maintainer is code-yeongyu, backed by Sisyphus Labs (sisyphuslabs.ai). One notable detail: the project uses an AI assistant called Jobdori (a custom fork of OpenClaw) to build and maintain itself in real time, streamed live on Discord as Building in Public -- you can watch it get assembled line by line by AI. The README has Japanese, Korean, Chinese, and Russian versions, signaling strong traction in East Asian communities.
2. Core selling points: three cards
Multi-Model Orchestration. The hardest differentiator. Instead of "pick one model and ride it to the end," it pulls multiple models into the same workflow and assigns tasks by strength. The README's Team Mode screenshot shows Kimi K3 and GPT-5.6 Sol in the same frame -- one handling reasoning, the other code generation, each doing its own part. This echoes the README's thesis: "The future isn't picking one winner; it's orchestrating them all. Models get cheaper every month." The logic: models iterate fast and prices drop monthly; locking into one vendor means betting the wrong direction, so build an orchestration layer that makes models pluggable.
Parallel Background Agents. It can run multiple agents simultaneously on different subtasks, without waiting for one to finish before starting the next. For example, a large task splits into "check docs," "fix tests," "write the API" -- three agents push forward in parallel and results are merged at the end. This is much faster than single-threaded serial execution, especially in large refactoring scenarios.
Crafted LSP/AST Tools. This is the precision card. Most coding agents rely on reading files plus regex matching to understand code; oh-my-openagent uses the Language Server Protocol (LSP) and Abstract Syntax Trees (AST) for structured context -- symbol definitions, reference relationships, type information, the same machinery your IDE uses when you jump to a definition. The result: more accurate edits, fewer accidental breakages.
3. How it divides turf with Claude Code, crush, and pi
Claude Code is Anthropic's official terminal/desktop agent: closed-source, locked to Anthropic models, subscription-based. oh-my-openagent's README calls it out directly: "Claude Code is a nice prison, but it's still a prison. You don't need to pay $200 for 2 hours of work." crush is Charm's open-source terminal agent, model-swappable, licensed FSL-1.1-MIT (see this site's crush resource). pi (earendil-works/pi) is another lightweight open-source agent (see this site's pi resource).
The fundamental difference with oh-my-openagent: it is an OpenCode plugin (parasitically enhancing a framework), not a standalone agent; it focuses on "multi-model in-frame orchestration," while crush and pi lean toward "single model, swappable"; its LSP/AST tools are crafted in-house, not just calling off-the-shelf ones.
| Dimension | oh-my-openagent | Claude Code | crush | pi |
|---|---|---|---|---|
| Positioning | OpenCode plugin, model orchestration | Terminal/desktop agent | Terminal agent | Lightweight agent |
| Models | Multi-vendor in-frame orchestration | Anthropic only | Any, swappable | Any, swappable |
| License | SUL-1.0 | Closed-source | FSL-1.1-MIT | Open-source |
| Cost | Software free, model self-paid | Subscription | Software free, model self-paid | Software free, model self-paid |
4. Getting started: install, wire models, fire up Team Mode
Install. oh-my-openagent is an npm package, most direct path:
npm install -g oh-my-opencodeAfter install, the command line has oh-my-opencode (or oh-my-openagent, omo-agent-toolkit). If you are a Codex user, take the LazyCodex path: npx lazycodex-ai install; the site at lazycodex.ai has instructions.
Wire models. Because it is multi-model orchestration, you can configure several providers at once. Set each vendor's API key (ANTHROPIC_API_KEY, OPENAI_API_KEY, MOONSHOT_API_KEY, etc.), then select the combination you want to orchestrate in Team Mode. The README's screenshot uses Kimi K3 + GPT-5.6 Sol; you can swap in a domestic combo like DeepSeek + Qwen to control cost.
Fire up Team Mode. The core mechanic is assigning different models and subtasks to different agents running in parallel. For exact configuration syntax, see the README and omo.dev docs -- the key mental model is "multi-model in-frame": not switching, but orchestrating.
5. The price of the pitfalls: license is SUL not MIT, plus a few more
Pitfall 1: the license is not MIT/Apache, it is SUL-1.0 (the one to read). GitHub's SPDX detector returns NOASSERTION for oh-my-openagent. Open the repo's LICENSE.md and the header reads "Sustainable Use License v1.0," abbreviated SUL-1.0, copyright Sisyphus Labs. SUL is not an OSI-approved open-source license; its core feature is "restricted use": you may use, modify, and build on it, but "competitive use" is restricted -- meaning you cannot take it to build a product or service that competes with oh-my-openagent itself. Third-party components retain their own licenses. This is a different license design from crush's FSL-1.1-MIT (which converts to MIT after two years) -- do not conflate the two. Personal coding or internal company use is generally fine; forking it into a competing SaaS requires reading the SUL text with a lawyer first.
Pitfall 2: it is an OpenCode plugin, not a standalone agent. Anyone coming in expecting "install and go" should note: oh-my-openagent depends on the OpenCode framework, so you need to understand how OpenCode works first. If you want a zero-setup standalone tool, a Claude Code subscription or crush is easier.
Pitfall 3: v5.0.0-beta.6, still beta. The current version number carries "beta," meaning interfaces and features are still shifting. Created 2025-12-03, pushed today, fast iteration is good, but before production integration, lock the version and run your scenario yourself.
Pitfall 4: model API is self-paid, and multi-model orchestration burns more. The software is free, but multi-model in-frame means multiple models burning tokens simultaneously. In Team Mode, Kimi + GPT running together means costs from both vendors stack up. Advice: cheap models for grunt work, strong models only for critical steps, do not run everything wide open the whole time.
Pitfall 5 (a sweet spot in disguise): anti-walled-garden, domestic-model friendly. The project's entire philosophy is "do not get locked into one vendor," so wiring domestic models is a design intent, not a compatibility shim. Kimi (Moonshot) is one of the primary models in the README's screenshot, and DeepSeek and Qwen are also supported. Compared to a closed-source agent locked to a single overseas model, oh-my-openagent is actually stronger for domestic use. For model choice, see this site's AI coding plan comparison.
Suited for: people already on OpenCode who want to add multi-model orchestration; people who refuse to lock into one AI vendor and want to freely combine models; anyone who needs parallel background agents for large refactors; Chinese developers who want to wire domestic models to control cost. Skip it if: you want a zero-setup standalone agent -- Claude Code or crush fits better; you want lightweight minimalism -- pi is the better pick.
Frequently Asked Questions
Q1: How is oh-my-openagent different from Claude Code and crush? A1: Claude Code is Anthropic's closed-source subscription agent, locked to Anthropic models; crush is a standalone terminal coding agent (FSL-1.1-MIT license, converts to MIT after two years); oh-my-openagent is an OpenCode plugin (not a standalone agent) focused on multi-model orchestration, licensed under SUL-1.0 (non-OSI). Different positioning and licensing for all three.
Q2: Can I use the SUL-1.0 license commercially? A2: Personal coding and internal company use are generally fine. But "competitive use" is restricted -- you cannot build a product or service competing with oh-my-openagent itself. Read LICENSE.md with a lawyer before building a rival SaaS. Note it has no "converts to MIT after two years" mechanism like crush's FSL.
Q3: Do I need to install OpenCode first? A3: Yes. oh-my-openagent is a "batteries-included" plugin for OpenCode, dependent on the OpenCode framework, not a zero-setup standalone tool. For a standalone agent, pick Claude Code or crush. Codex users can use LazyCodex (npx lazycodex-ai install).
Q4: How do I control costs with multi-model orchestration? A4: The software is free, but multi-model in-frame means multiple models burn tokens simultaneously, stacking costs. Use cheap models for grunt work, strong models only for critical steps, and do not run everything wide open. Domestic models (Kimi/DeepSeek/Qwen) cost less and wiring them is a design intent, not a shim.
Q5: It is still beta -- can I use it in production? A5: Currently v5.0.0-beta.6; interfaces and features are still shifting with fast iteration. Before production integration, lock the version and run your scenario yourself. Created 2025-12-03, pushed today, highly active, but stability needs your own verification.
References
- code-yeongyu/oh-my-openagent GitHub repo (67,728 stars / 5,538 forks, TypeScript, SUL-1.0 license): https://github.com/code-yeongyu/oh-my-openagent
- oh-my-openagent official site (omo.dev): https://omo.dev
- LICENSE.md (SUL-1.0 full text, including use-restriction clauses): https://github.com/code-yeongyu/oh-my-openagent/blob/main/LICENSE.md
- npm package oh-my-opencode (v5.0.0-beta.6): https://www.npmjs.com/package/oh-my-opencode
- LazyCodex (OmO for Codex install path, site lazycodex.ai): https://lazycodex.ai
- This site: crush resource | pi resource | openclaw resource | AI coding plan comparison