Your Claude Subscription Works in OpenCode Again — Through a Door Anthropic Has Not Confirmed

One-line news: your Claude subscription works in OpenCode again

On 30 October 2026, the open-source coding agent OpenCode shipped v2.2.3, and its most talked-about change fits in one sentence: install Claude Code, run claude auth login once, then select Claude Agent in /connect to use your own Claude Code login and models. The sentence deserves a full article because of the door it reopens. In January 2026, Anthropic tightened restrictions on third-party harnesses using Claude subscriptions; in March, OpenCode removed its built-in Anthropic login plugin. Since then, subscription users were locked out of OpenCode — API keys or workarounds were the only routes. On 30 October the door reopened, but through a different passage: Anthropic's own Claude Agent SDK, not a revival of the old plugin. How the door reopened, and how long it stays open, is exactly what needs explaining.

First, get the steps right, because several garbled versions are already circulating. Per the v2.2.3 release notes, the flow is: install Claude Code on your machine; run claude auth login; then select Claude Agent in /connect — not "Anthropic". Picking the right entry matters: the Anthropic entry is the old API-key route, while Claude Agent is the new built-in provider. If your claude binary is not on the default path, the release notes offer an OPENCODE_CLAUDE_PATH environment variable to point at it. Throughout the flow, OpenCode never touches your saved login credentials; signing in happens inside your own Claude Code. That design detail is the heart of the legitimacy debate, and we return to it below.

Why 30 October matters: the January crackdown and the March removal

To weigh this update, the year's timeline has to come first. According to OpenCode community records, on 9 January 2026 Anthropic tightened safeguards against third-party harnesses spoofing the Claude Code harness, after accounts were banned for triggering abuse filters from third-party harnesses using Claude subscriptions. The same day, an OpenCode issue titled "Broken Claude Max" appeared. On 19 March, OpenCode merged a pull request bluntly titled "anthropic legal requests", removing the built-in opencode-anthropic-auth plugin and dropping Anthropic from the provider list. From then on, the subscription-login route inside OpenCode was dead. The community workarounds that followed were exactly the kind of thing January's crackdown targeted: plugins that read the saved login file or pulled tokens out of the OS keychain and replayed them while pretending to be Claude Code. They worked, nervously, with the account risk on the user.

The new provider is designed against that history. Its idea in one sentence: instead of copying, storing or replaying your login, run the claude program already installed on your machine through Anthropic's own Claude Agent SDK. Per the provider spec, OpenCode "never sits between Claude Code and Anthropic or changes Claude Code's requests". That distinction is not wordplay: the old plugin took your credentials and used them itself; the new provider lets the official program use them itself. There is also a neat tool swap: Claude Code's built-in Bash and Edit tools are switched off and replaced with OpenCode's own tools, so every file edit and shell command goes through OpenCode's permission system rather than around it. Each OpenCode conversation maps to one Claude Code session and resumes it on every message, so context survives the change of shell.

What else shipped in v2.2.3

Focusing only on Claude Agent would miss two other worthwhile changes. First, desktop long-session performance: streaming replies and long conversations now do much less rendering work, alongside fixes for timeline growth, zoom and DOM-observer performance — for users who keep one session open from morning to night, this may matter more in daily feel than the new provider. Second, desktop stability fixes: side-panel tabs, repeated "Session not found" toasts, and "Open file" over HTTP in the web app — several of them regressions introduced in 2.2.2, now cleaned up.

Two breaking changes need attention before upgrading. TUI plugins now require OpenTUI 0.8 or later: in OpenTUI 0.8, a ScrollBox's wrapper, viewport, content and scrollbars became facade objects instead of renderables, and a custom RenderContext must implement a clock and an on-animation-tick callback — old plugins will break without an upgrade. And managed policy got split: integration.use now covers integration accounts only, while new mcp.use, skill.use and plugin.use actions govern MCP servers, skills and plugins. Rules with the old mcp:, skill: and plugin: prefixes convert automatically, but a bare integration.use rule no longer covers MCP servers, skills or plugins. Admins who previously relied on one rule to cover everything must add the three new ones, or they will have a gap they believe is closed.

The central question: did Anthropic approve it?

Now the sensitive point: has Anthropic approved this provider? As of the checks on 1 November, the public record shows no statement from Anthropic on it. On one side, OpenCode says it has Anthropic's permission. On the other, Anthropic's Agent SDK documentation still carries a restriction, in substance: unless previously approved, third-party developers may not offer claude.ai login or rate limits for their products, including agents built on the Claude Agent SDK. That restriction has been quoted verbatim by several trackers of the story. Put together, the picture is delicate: technically the implementation is much cleaner than the January-era workarounds — official SDK, no credential handling — but on the rules text, whether the "previously approved" condition is met is not publicly evidenced.

What that uncertainty means depends on who you are. For paying Pro or Max subscribers, the short-term benefit is real: no separate API spend is needed to run Claude models in OpenCode, and the cost display reads $0 on a subscription because the subscription covers usage. But the risk is equally real: if Anthropic decides this route is outside the approved scope, January and March are the precedent for how fast a route can close — and account standing could be involved. For API-key or cloud-account users, the impact is far smaller: the new provider supports both logins, costs show at list price, and the rules dispute barely touches you. For team admins, the prudent move is a small pilot, not a fleet-wide switch of the default provider, until Anthropic says something definite.

One more signal is easy to miss: later the same day, v2.2.4 followed with a guided setup flow that helps copy the login command, checks the installation and verifies the connection — seven hours from "the feature exists" to "a button installs what it needs". That pace suggests the OpenCode team expected onboarding friction and wanted "it works" to become "it works easily". It also suggests the 2.2.3 first-hour experience was rough. If you are upgrading on the news, go straight to 2.2.4 or later rather than fighting setup issues on 2.2.3.

How this differs from our earlier OpenCode coverage

We have previously covered OpenCode 2.0 and the OpenCode Console API-key leak. This story overlaps with neither. Version 2.0 was about the product growing up — background services, code mode, capability building. The leak story was about trust in a hosted service: a cloud-side security lesson. This one is about access: the subscription-login route reopening, and the rules contest behind it. Together the three trace OpenCode's second half of 2026 along three lines — capability, trust, access. The subscription question belongs to the access line, and its outcome depends less on how well the code is written than on how Anthropic's rules are interpreted. When deciding, read both sides: the feature list in the release notes, and the risk picture in the rules text and history. Neither alone is enough.

One claim already circulating should be corrected: that "OpenCode has received official authorisation from Anthropic". That wording is not currently supportable. The release notes describe functionality and usage and attach no authorisation document, and OpenCode's "we have permission" statement has no public Anthropic confirmation behind it so far. In this article we keep "has permission" strictly attributed to OpenCode, and "no public statement" as a check result, not a guess. The distinction matters: if you recommend the route to a team on the basis of a rumour treated as fact, and the rules tighten, the responsibility lands on the recommender.

Advice for three kinds of reader

If you are an individual subscriber and want to try it today: confirm a current Claude Code install that signs in cleanly, upgrade OpenCode to 2.2.4 or later, pick Claude Agent in /connect (not Anthropic), choose a Claude model under /models, and verify with one small task before migrating working sessions. Check whether OPENCODE_CLAUDE_PATH needs setting if your claude binary lives off the default path — setting it early saves a round of debugging.

If you run a team: do not touch the fleet default yet. Pilot on one or two non-critical projects and verify three things: that the permission system really takes over file and shell operations as designed; that the split managed-policy rules for MCP, skills and plugins are all configured to your requirements; and that subscription usage really displays as $0 without anyone mistaking "displays $0" for "unlimited". Keep January and March in the internal assessment: this route has been tightened once before, so keep a rollback path — at minimum the API-key route — ready.

If you are still watching: the thing to track is not "can I use my subscription for free today" but where the rules go. Two signals are enough: whether Anthropic issues any public statement on the Claude Agent provider, and whether later OpenCode versions adjust the implementation (for example, further restricting login methods). Either signal will tell you more about how long this route lasts than star counts or trending posts. Until then, treat it as an option that works but carries uncertainty — that is the honest framing.

Three login routes, three bills

The new provider accepts three logins, and their billing differs — worth separating before month-end surprises. Subscription login, the star of this release, shows $0 in OpenCode because the subscription already covers usage; the cost is the fixed monthly fee, and rate limits and usage policy still apply. API-key login shows the models' list prices, billed on Anthropic's API side — a separate ledger from the subscription. Cloud-account login likewise shows list prices. Three routes share one entry point but different billing homes, so teams mixing subscribers and key users must reconcile by login type and never read "displays $0" as "spent nothing".

One more practical note belongs in any team handbook: a $0 display is not a cost of zero, and it is not unlimited either. Past account bans for abuse-filter triggers from third-party subscription use are a reminder that a subscription is not an unlimited pass. The new provider no longer copies credentials, but the rhythm and concurrency of its requests can still look anomalous to the server side. Start small, watch for rate-limit behaviour over a few weeks, then migrate real workloads — and keep the API-key route as a fallback so a subscription-channel interruption does not stop all work.

Three rule changes side by side

Lining up the year's three changes clarifies the trajectory. On 9 January, the change came from Anthropic's side: enforcement tightened against harnesses spoofing Claude Code, after subscription abuse-filter bans. On 19 March, the change came from OpenCode's side: the "anthropic legal requests" merge removed the built-in login plugin and the Anthropic provider entry — a compliance move that traded a contested entry point for continued normal operation. On 30 October, the change is in implementation: no credential handling, official SDK, the official program signing in itself. The rules text never publicly changed; what changed, each time, is the understanding of what counts as compliant. January taught "don't pretend to be Claude Code"; March taught "don't hold other people's credentials"; October's attempt is "let the official program sign in, and I only take over its tools". Whether that attempt passes depends on Anthropic's reading of "previously approved" — not on how elegant the code is. In rules questions, code is not evidence; a statement is.

For the next two to four weeks, watch three things. First, any public response from Anthropic on the Claude Agent provider, even a one-line acknowledgement. Second, whether later OpenCode versions narrow the implementation — implementation tightening is often the early sign of rules pressure. Third, whether reports of subscription accounts being restricted cluster in the same period; a single case proves little, but a cluster deserves attention. If none of the three moves, the route is at least running steadily for now — though that is not the same as risk-free. If any of them moves, it is time to re-evaluate the subscription channel.

To close in one sentence: this is an update worth trying, but worth trying small first — the feature is real, and so is the uncertainty, and both belong in your decision.