Bring back Anthropic auth to OpenCode. Pick your path — Claude Code session, browser OAuth, or plain API key.

OpenCode 1.3.x removed built-in Anthropic auth. This plugin restores it as a standalone package with two login options and a config helper that sets up the Anthropic provider for you.

Both paths are offered during opencode auth login -p anthropic. You choose once and the plugin remembers.

Install the opencode-claude-code-auth plugin and configure Anthropic models in this project's opencode.json by following: https://raw.githubusercontent.com/ragingstar2063/opencode-claude-code-auth/main/README.md 

The plugin hooks into OpenCode's provider system and does three things:

1. Intercepts Anthropic requests. It registers an auth.loader with a custom fetch that takes over all traffic to the Anthropic API. Every outgoing request gets rewritten with the right auth headers, beta flags, tool name format, and system prompt identity before it leaves. Every response stream gets the tool names translated back.

2. Manages tokens. Depending on which login path you chose:

• Claude Code session — reads tokens from the macOS Keychain (or ~/.claude/.credentials.json on Linux/Windows), refreshes them by spawning claude with stdin closed when they're about to expire, and re-reads the credentials file afterward regardless of exit code.
• Browser sign-in — runs a PKCE OAuth flow against Anthropic's authorize endpoint, exchanges the code for tokens via application/x-www-form-urlencoded POST, and refreshes them via the token endpoint when they expire.
• API key — the plugin steps aside and lets OpenCode handle it natively.

Both OAuth paths sync tokens to OpenCode's auth.json via client.auth.set. On 401/403 responses, the plugin forces a refresh and retries once.

3. Patches your config. On clean OpenCode installs, Anthropic models don't appear until the provider is defined. The plugin can add the Anthropic provider block to opencode.json — either through a CLI prompt during login, or via npx opencode-claude-code-auth configure.

Claude Code session path:

Browser sign-in path: tokens are stored in OpenCode's auth storage after the initial OAuth exchange.

API key path: managed by OpenCode directly.

Any model your Anthropic subscription supports will work. The default config adds:

Add more by editing the provider.anthropic.models block in opencode.json.

OpenCode 1.3.x needs explicit Anthropic provider config. The plugin handles this two ways:

During login — if Anthropic isn't configured, the CLI asks:

From the command line:

npx opencode-claude-code-auth configure # interactive npx opencode-claude-code-auth configure -y # skip confirmation npx opencode-claude-code-auth configure --global # global config npx opencode-claude-code-auth doctor # check what's configured

{ "$schema": "https://opencode.ai/config.json", "plugin": ["opencode-claude-code-auth"], "model": "anthropic/claude-sonnet-4-5", "small_model": "anthropic/claude-haiku-4-5", "provider": { "anthropic": { "npm": "@ai-sdk/anthropic", "name": "Anthropic", "models": { "claude-sonnet-4-5": { "name": "Claude Sonnet 4.5" }, "claude-haiku-4-5": { "name": "Claude Haiku 4.5" } } } } }

The Anthropic OAuth API expects requests shaped differently from standard API-key requests. The plugin handles this transparently:

Quick fix: Run npx opencode-claude-code-auth configure then opencode auth login -p anthropic.

Full guide: docs/troubleshooting.md

Defaults work for most setups. If Anthropic changes something before an update ships, set an env var:

• Setup — all install and configure options
• Auth Methods — how each path works, refresh behavior, decision guide
• Troubleshooting — every error and its fix
• Architecture — module map, request flow, design constraints

• opencode-anthropic-auth — original built-in Anthropic auth (deprecated)

MIT. See LICENSE.