Ten months ago, Anthropic nailed the door shut on third-party coding tools borrowing a Claude subscription. On October 30, OpenCode quietly unscrewed the hinges. The release note is one bullet buried between desktop rendering fixes and a plugin migration warning: install Claude Code, run claude auth login, then pick Claude Agent in /connect. No press release. No blog post. No triumphant thread from Anthropic explaining what changed. Just a new provider entry that lets a Claude Pro or Max plan walk back into the exact harness Anthropic spent the winter legally escorting out of the building.
The internet noticed anyway. OpenCode's announcement, essentially the words you can now use your claude subscription in OpenCode, cleared more than a million views on X. Developers reacted the way people react when an ex shows up at the same party, acts normal, and orders your favourite drink.
The punchline was never technical.
What Actually Shipped
The built-in Claude Agent provider landed in PR #58782, contributed by Tarquinen, and it is almost insultingly simple. OpenCode does not impersonate anything or recycle OAuth tokens. It runs the real Claude Code binary you installed, on your machine, and lets Anthropic's own tool make the model calls. OPENCODE_CLAUDE_PATH exists if you keep Claude in an unusual corner of your disk.
The setup ritual is four steps and roughly one coffee:
- Install Claude Code if you have not already, because OpenCode refuses to pretend it is the model layer.
- Run
claude auth loginin your terminal and finish the browser sign-in. - Open
/connectinside OpenCode and choose Claude Agent from the provider picker. - Pick your model and get back to arguing with a diff at 1am, which is the actual hobby here.
Version 2.2.4 turned that ritual into a guided flow that copies the login command, checks the install, verifies the connection and recovers from the Needs sign-in state. It also replaced Claude Code's own misleading /login hint with claude auth login, the software equivalent of correcting your friend's grammar mid-apology. As of November 13 the project is on 2.3.0.
The Breakup, Briefly, For Context
January 9, 2026 is the date that started it. Anthropic deployed server-side checks that blocked tools like OpenCode, Cline and RooCode from authenticating to Claude Pro and Max through OAuth, because those tools had been sending headers that made requests look like they came from the official Claude Code client. The commercial logic was never mysterious: a subscription subsidises first-party use, and routing agent workloads through it turns a subsidy into a bill.
The execution is what lit the fuse. No advance notice. Abuse-filter bans for some accounts. A cutover at 02:20 UTC, right in the middle of somebody's workflow. OpenCode shipped ChatGPT Plus support within hours, then watched Anthropic formalise the rule in its Terms of Service on February 19. On March 19 came the legal requests and PR #18186, commit message anthropic legal requests, deleting every reference to Claude Pro and Max authentication from the codebase. April 4 extended enforcement to all third-party harnesses, OpenClaw and NanoClaw included.
- January 9: subscription OAuth cut off for third-party harnesses.
- March 19: OpenCode removes Claude Pro and Max login after legal requests.
- April 4: enforcement widens to every third-party harness.
- October 30: Claude Agent provider ships, login walks back in.
Star counts tell the story of what that fight cost. OpenCode now sits at roughly 226,000 GitHub stars against about 155,000 for Anthropic's own claude-code repository. Some of that gap is genuine adoption. Some of it is people starring a repository as a protest vote, which is the cheapest form of rebellion ever invented.
Why this version is legally cute. The new route dodges the original complaint almost completely. Claude Code makes the model calls through Anthropic's Claude Agent SDK. OpenCode supplies its tools, permissions and subagents. OpenCode never reads or stores your credentials, which means the January objection, credential recycling through a lookalike client, simply is not happening.
Is Anthropic on board? Officially, nobody has said so. The SDK documentation still states that Claude.ai logins inside third-party products require prior approval. Since June, programmatic use on paid plans draws on a fixed monthly Agent SDK credit. Amp shipped a similar Claude-plan integration in October, so this looks less like a loophole and more like a quiet thaw. Theo Browne summed it up as nice to see Anthropic loosening the grip on the sub a bit.
What It Means If You Code For A Living
Two coding tracks now run in parallel and both got stronger this autumn. Anthropic keeps the managed track sticky with deeper orchestration and more capacity. OpenCode keeps the neutral track cheap, because a harness that can swap providers cannot be held hostage by a billing page.
The practical read is simple. If you already pay for Claude Max and you prefer OpenCode's terminal, you no longer choose between them. If you were paying API rates purely to keep OpenCode alive, your bill has a new ceiling. If you run a team, the compliance question just got harder, because the same login is now reachable from two harnesses and your audit trail only sees one of them.
- Watch whether Anthropic documents the arrangement or leaves it as an approved-by-silence feature.
- Watch the Agent SDK credit, since a fixed monthly allowance is where plan economics quietly get renegotiated.
- Watch whether Cline and RooCode get the same path, because that is when this stops being an OpenCode story.
The funniest part is how little ceremony the reunion got. Ten months of legal letters, a Terms of Service rewrite, a codebase purge and 226,000 stars later, the fix was a provider dropdown and an honest handshake with the official binary. Anthropic never announced it. OpenCode never explained it. The changelog just says you can use your Claude Code login now, and somewhere a very tired developer reopened a tab they closed in January.
Comments