An Eight-Month-Old Install Guide: oh-my-opencode Isn't Called That Anymore

I was looking for an alternative to Claude Code. Two conditions: open source, and able to run local LLMs. That’s how I ended up saving a write-up covering the installation and configuration of oh-my-opencode, an agent plugin that sits on top of OpenCode. Install command, authentication, config file layout, per-agent model assignment — all in one sweep, which looked like a good starting point.

It’s from January 7, 2026, and it’s now September. Eight months. That’s a long time for an install guide, and it shows. Checking it, the project has been renamed, and what the write-up warned against as “not recommended” was formally enforced in April. There are also a few places where the article contradicts itself.

Table of contents

The name changed first

Go to github.com/code-yeongyu/oh-my-opencode and you land on oh-my-openagent. The official docs give the reason.

The project was formerly called “oh-my-opencode” but is now “oh-my-openagent” to reflect its provider-agnostic design.

A rename meant to shake off the impression of being OpenCode-only. There’s now a separate lightweight edition for Codex CLI (npx lazycodex-ai install) too.

Eight months, side by side:

The agent roster didn’t just grow, it changed character. The article’s Sisyphus, Oracle, Librarian and Explore are still there, joined by Hephaestus (deep autonomous work), Prometheus (strategic planning), Atlas (todo management), Metis (gap identification), Momus (plan review), and Sisyphus-Junior.

Once the install command stops working, you can’t trust the rest of the guide either. Config paths, flags, model names — all of it assumes that name.

What the article warned about actually happened

The section the article invests the most in is Anthropic authentication. Two days after publication, an update block was added.

[2026-01-09 update] Anthropic OAuth authentication restriction

Anthropic has applied a technical restriction making Claude Code OAuth tokens usable only from official Claude Code.

And its conclusion:

Since this isn’t a simple block but has produced account bans, be sure to use the API key method.

That call was right. And it went one step further afterwards. Per an April 3, 2026 report, Anthropic cut off using Claude Pro/Max subscriptions with third-party agents entirely, effective April 4 at 12pm PT. Boris Cherny, who heads Claude Code, put it this way:

We’ve been working hard to meet the increase in demand for Claude, and our subscriptions weren’t built for the usage patterns of these third-party tools.

Not a technical sanction so much as a policy decision driven by capacity. Existing subscribers were offered a one-time credit equal to a month’s plan price (redeemable until April 17) and up to 30% off pre-purchased usage bundles, with pay-as-you-go or the API as the remaining paths.

So the article’s “use an API key” advice stands — for a different reason. The article cited ToS violation and account bans; what actually settled was a policy removing third-party use from the subscription products. For someone reading it today, “that route doesn’t exist” is more accurate than “it’s a violation, so avoid it.”

Worth noting: the oh-my-openagent repo carries a banner reading “Anthropic blocked OpenCode because of us.” That’s where it places itself in this story.

Places the article contradicts itself

Separate from what aged, some passages were already self-contradictory at publication.

Both star figures are labeled “as of early January 2026.” It looks like the intro was refreshed later and the body table wasn’t.

The -preview one will actually stop you mid-setup. Partway through, the article says GA shipped so the suffix is unnecessary and recommends google/gemini-3-flash:

"model": "google/gemini-3-flash"
"model": "google/gemini-3-pro"

Troubleshooting item 5 in the same article names exactly that spelling as the cause of a 404.

// ❌ wrong (produces 404)
"model": "google/gemini-3-flash"

// ✅ correct
"model": "google/gemini-3-flash-preview"

A head-on collision. And the article’s own config example (oh-my-opencode.json) uses -preview, which suggests the GA note was slipped in later without updating the rest. Either answer might be right, but when one article gives two, the reader can’t choose.

Same with the agent count. “7 specialist agents led by the Sisyphus orchestrator” makes 8 total, while the table says 7 — and counting the table’s own list gives 7 including Sisyphus.

Three different repositories

The GitHub issue links the article cites point at three different organizations.

• anomalyco/opencode/issues/6930 — the Anthropic OAuth item
• opencode-ai/opencode/issues/{523,530,541} — exit and background issues
• sst/opencode — listed under references as “OpenCode source code”

Checking it: the repository the official docs (opencode.ai/docs) point at is anomalyco/opencode, and the site footer carries Anomaly’s copyright. sst/opencode is the earlier org name.

The middle one is the problem. opencode-ai/opencode is an archived, different project. Its banner says:

This repository is no longer maintained and has been archived for provenance. The project has continued under the name Crush, developed by the original author and the Charm team.

A separate Go-based terminal AI tool, now continued as Charm’s Crush. The names collided, apparently — which means the supporting links for troubleshooting items 7 and 8 point at the wrong project. The symptom descriptions are still useful; you just can’t follow the link to verify them.

The Claude Code comparison doesn’t hold up now

The article’s FAQ says:

Claude Code can only use Claude models, whereas oh-my-opencode uses Claude, ChatGPT and Gemini optimized per role. It also provides multi-agent collaboration, parallel execution, Todo Enforcer and other features Claude Code doesn’t have.

The first half is fair. The second half doesn’t match the current documentation. From Claude Code’s subagents page:

Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions.

For independent investigations, spawn multiple subagents to work simultaneously.

Background subagents run concurrently while you continue working.

Multi-agent and parallel execution both exist. Same for the comparison table’s “model selection: single model only.”

The model field controls which model the subagent uses.

Each subagent gets its own model. Within the Claude family, yes — but not “single model.”

The hook names don’t line up either. The article claims “fully compatible with Claude Code” and lists PreToolExecution / PostToolExecution / Notification / Stop. Claude Code’s documented event names run to more than thirty and include PreToolUse and PostToolUse. Notification and Stop overlap; the first two have different names. From the article’s list alone there’s no way to check what “fully compatible” means.

That said, this is the item to weigh against the eight-month gap. Comparison pieces spoil first by nature. The point is that it doesn’t hold for someone choosing a tool today, not that the author invented anything.

The slash commands the title promises

The title advertises “basic commands, slash commands, integration methods and more.” In the body there are three slash commands, and all three pass by in troubleshooting or footnote context.

• /connect — OpenAI OAuth login
• /exit — quitting (while explaining that background agents survive it)
• /tasks — checking running tasks

No /init, no /model. There’s no basic-commands chapter either, because the author says up front to read Part 1 first. It’s the second entry in a series, so this is less a flaw in the article than a title that over-promises to anyone who lands here from search.

What still holds

Aside from what aged, some of the design is unchanged eight months on.

Automatic context injection. Read a file and every AGENTS.md and README.md from that directory up to the project root gets injected, with each directory’s context injected once per session. The current docs still carry this.

Curated MCPs. Exa (web search), Context7 (official docs), and Grep.app (GitHub code search) bundled by default — also still in the current docs.

Multi-account load balancing. Registering several Google accounts to get past rate limits is still there. Though the article introducing it as “getting past rate limits for free” and then immediately adding “but you must comply with Google’s terms of service” doesn’t add up. Recommend the workaround or require compliance; together they leave the reader nothing to judge with.

The warnings about cost. Paying for a subscription and API usage both, the price of Opus-class models, the advice to use the stock tool for simple work — all still valid. More valid since April, if anything.

The local LLM side is still empty

This is where my reason for saving it runs aground. The article’s FAQ answers:

A: oh-my-opencode is designed around Claude, ChatGPT and Gemini subscriptions. Compatibility with local models isn’t clearly addressed in the official documentation, and would need actual testing.

An honest answer, and eight months later not much has changed. Where Ollama appears in the current docs is one line of Sisyphus’s fallback chain, and the name there is ollama-cloud — which reads as Ollama’s hosted service rather than local execution. There’s a mention of detecting custom-endpoint providers during setup, but no guidance on wiring up a local model that I could find.

The recommended configuration says the same thing. The model the docs suggest for Sisyphus is Claude Opus 5, and the rest of the chain is hosted models like Kimi K3, GPT-5.6 and GLM-5.2. The rename to oh-my-openagent was for being “provider-agnostic” — and provider here means several companies, not locally.

Wrapping up

• The project is now oh-my-openagent. The install command is bunx oh-my-openagent install, the minimum OpenCode version went from 1.0.150 to 1.4.0, and the agent count went from 7 to 11.
• The Anthropic subscription cutoff the article warned about took effect on April 4, 2026 — as a policy removing third-party use from the subscription products, not as a ToS enforcement action.
• The article contradicts itself in five places. The Gemini 3 -preview one in particular will stop you mid-setup, since the body and the troubleshooting section say opposite things.
• The cited repositories span three organizations. One of them, opencode-ai/opencode, is an archived, different project (now Crush), so those supporting links point somewhere else entirely.
• The Claude Code comparison doesn’t hold today. Subagents, parallel execution and per-agent model selection are all documented.
• The slash commands the title promises amount to /connect, /exit and /tasks. It’s an article written on top of Part 1.
• Still valid: context injection, curated MCPs, multi-account balancing, and the cost warnings.
• Local LLMs still have no guidance. The Ollama that appears in the docs is ollama-cloud, and every recommended model is hosted.

In a tool write-up the first things to spoil are the install command and the comparison table. Both are facts about a moment. What lasted was why it’s built that way — how context gets injected, what gets delegated, where the cost leaks.

Back to the two conditions I was looking for, and they split. Open source: still yes. Local LLM: still no. If anything, April’s subscription cutoff made this tool’s position clearer — it splits hosted models from several companies across roles; it doesn’t run on your machine. For not being tied to Claude Code alone, it fits. For not calling out to anything at all, it isn’t the one.

Opening a saved install guide eight months later will keep happening. Open the repository first is the ordering I took from this. Checking whether the name is still the name decides whether the rest is worth reading.

References

• code-yeongyu/oh-my-openagent — GitHub
• Oh My OpenAgent Docs
• OpenCode Docs
• opencode-ai/opencode (archived) — GitHub
• Anthropic cuts off the ability to use Claude subscriptions with OpenClaw and third-party AI agents — VentureBeat
• Subagents — Claude Code
• Hooks — Claude Code
• Source: Open Code review (2): installing and configuring oh-my-opencode