Yesterday I published a blog post about how I use OpenCode daily and don’t miss Claude Code. I documented my entire workflow — custom commands, AGENTS.md, MCP servers, the /connect flow with my Claude Max subscription. I ended with “I don’t miss Claude Code at all.”
Today I’m writing about switching back.
On March 19, Dax — one of OpenCode’s core maintainers — merged PR #18186 with a two-word commit message: “anthropic legal requests.” The PR removed the Anthropic OAuth plugin, the Claude system prompt, and every reference to Claude Pro/Max authentication from the codebase. 437 developers hit the thumbs-down reaction. The section in my article where I walked through /connect → Anthropic → Claude Pro/Max? That flow no longer exists.
Anthropic didn’t just block OpenCode technically. They sent lawyers.
The full timeline
This didn’t happen overnight. It was a slow escalation over three months.
January 9, 2026 — Anthropic silently deploys server-side checks. Third-party tools using Claude Pro/Max OAuth tokens start getting a blunt error message: “This credential is only authorized for use with Claude Code and cannot be used for other API requests.” No warning, no announcement. Tools like OpenCode, Cline, and RooCode stop working overnight.
January 15 — George Hotz publishes a blog post titled “Anthropic is making a huge mistake.” His prediction: “You will not convert people back to Claude Code, you will convert people to other model providers.”
February 19 — Anthropic formalizes everything into updated Terms of Service. A new “Authentication and credential use” section explicitly bans subscription OAuth tokens in any third-party tool, including the Agent SDK. The same day, OpenCode removes all Claude OAuth code from the codebase.
February 20 — The Register reports on the ban. DHH weighs in: “Terrible policy for a company built on training models on our code, our writing, our everything.”
March 19 — PR #18186 lands. The last traces of Anthropic integration are gone from OpenCode. The removal includes the built-in auth plugin, provider login UI hints, and documentation references.
What was actually happening under the hood
The technical reality is worth understanding. OpenCode and other third-party tools were authenticating users through the same OAuth flow that Claude Code uses. When you ran /connect in OpenCode, it opened your browser, you logged into your Anthropic account, and the tool received OAuth tokens — the same way Claude Code does.
Anthropic’s argument: these tools were “spoofing” the Claude Code client identity by sending headers that made Anthropic’s servers think requests came from the official tool. Anthropic implemented client fingerprinting to detect and block requests not originating from official clients.
The counterargument from developers: users were authenticating with their own paid subscriptions. They weren’t stealing anything. They were choosing which terminal interface to use with models they’re paying for.
The personal cost
I built my entire AI coding workflow around OpenCode. Custom commands for PR reviews, code analysis, type checking. A review agent with specific permissions and temperature settings. MCP server configurations. Project-specific AGENTS.md files committed to every repository. All documented, all version-controlled, all designed to be model-agnostic — which was the whole point.
The /connect flow with Claude Pro/Max was the zero-friction entry point that made the switch from Claude Code painless. Same subscription, same models, better tooling. That was the pitch, and it was true.
Now that path is gone. I can still use OpenCode with an Anthropic API key (billed per token), but the economics are completely different. The Max subscription gave me predictable, unlimited usage. API billing means watching token counts and second-guessing whether that deep reasoning task is worth the cost.
The alternative: switch to OpenAI models within OpenCode. OpenAI has done the opposite of Anthropic here — they’ve officially partnered with OpenCode, allowing Codex and ChatGPT subscribers to use their subscriptions directly in OpenCode, OpenHands, and RooCode. OpenAI is welcoming third-party tools while Anthropic sends legal requests.
But I prefer Claude’s models for coding. Sonnet and Opus are genuinely better at understanding complex codebases and producing clean, idiomatic code. That’s not loyalty — it’s been my experience across hundreds of sessions.
So here I am, switching back to Claude Code.
Did Anthropic kill OpenCode?
No. OpenCode has 112,000+ GitHub stars, 2.5 million monthly active developers, 800+ contributors, and a release cadence that puts most commercial software to shame. It doubled its stars since January. It’s not going anywhere.
But Anthropic dealt a real blow to OpenCode’s value proposition for a specific — and large — segment of users: people who were using it because they could bring their Claude subscription. The pitch was “same models, more freedom.” Now it’s “different models, more freedom” — which is a harder sell when you believe Claude is the best model for your work.
The community responded immediately. Within days of PR #18186, developers published workaround plugins — opencode-claude-auth, opencode-anthropic-oauth, routing proxies. The open source community is nothing if not resourceful. But workarounds that violate ToS are a ticking clock, not a solution.
Will this slow adoption?
The developer community is splitting along predictable lines.
Group 1: Stay with OpenCode, switch models. These developers value tool freedom over model preference. With the OpenAI partnership, they get subscription-based access to GPT-4o, o3, and Codex models directly in OpenCode. Some will discover they don’t miss Claude as much as they thought.
Group 2: Go back to Claude Code. These developers — I’m in this group, reluctantly — prefer Claude’s models enough that they’ll accept the vendor lock-in. Claude Code has improved significantly. It’s not the same tool it was when I left.
Group 3: Move to API billing. Some will keep using OpenCode with Anthropic API keys, paying per token instead of a flat subscription. This works for lighter usage but breaks down for power users who were running agents in loops overnight.
The question is which group is biggest. My guess: Anthropic is underestimating Group 1. George Hotz’s prediction feels right. When you tell developers they can’t use a tool they’ve chosen, some come back to your tool. More go find someone else entirely.
The 437 thumbs-down reactions on PR #18186 aren’t just frustration. They’re a signal. Every one of those developers is now making a choice, and “go back to the tool that just forced this situation” isn’t the obvious answer.
The irony
Anthropic’s models were trained on open-source code from the internet — code written by the same developers who build and use tools like OpenCode. The company built its business on the collective output of the open-source community, then blocked the open-source community’s tools from accessing its models through paid subscriptions.
Meanwhile, OpenAI — not historically the open-source champion — is actively partnering with third-party tools and extending subscription access. The roles have reversed in a way nobody predicted a year ago.
What I’m doing now
I’m moving back to Claude Code. I’ve already started migrating my custom commands to Claude Code’s slash commands (different format, same concept). My AGENTS.md files work with both tools — that’s the one piece that transfers cleanly. The MCP server configurations need rebuilding but are conceptually identical.
Claude Code has gotten better. The permission system, the plan mode, the tool execution — it’s more refined than when I left. I’m not going to pretend otherwise.
But I’m doing this under protest. Not because Claude Code is bad, but because the reason I’m switching isn’t technical merit — it’s a legal letter. I chose OpenCode because it was better for my workflow. I’m leaving because Anthropic decided I’m not allowed to use their models the way I want, even though I’m paying for them.
The real question
OpenCode will survive this. It has momentum, community, and now a partnership with OpenAI that gives it a clear path forward without Anthropic.
The real question is whether Anthropic’s walled garden strategy works. In the short term, they’ll recapture users like me who prefer Claude’s models. In the long term, they’re training developers to build workflows that don’t depend on Claude at all — which is exactly what OpenCode was designed to enable.
Every developer who switches to GPT-4o or Gemini inside OpenCode because of this ban is a developer Anthropic may never get back. Not because the models aren’t good, but because the trust is gone.
I hope Anthropic proves me wrong. I hope they open up subscription access to third-party tools and compete on product quality rather than legal enforcement. Claude’s models deserve to be used everywhere. The walled garden doesn’t serve them well.
Until then, I’ll be in Claude Code. It’s good software. I just wish I was here by choice.