Use your Claude Pro or Max subscription with OpenCode. Supports multiple accounts with automatic rotation when you hit rate limits.

AnomalyCo originally published this as a built-in plugin shipped with OpenCode itself. The repo was archived following a legal request from Anthropic. Before it was archived, rmk40 — a frequent contributor unrelated to AnomalyCo — had developed multi-account support that never made it upstream.

This fork continues that work and extends it with deeper research into how the official Claude Code CLI identifies itself to the API. The mimicry was derived by analyzing Claude Code's open source code on GitHub. This isn't just a system prompt injection — see the full mimicry analysis for details.

Educational purpose. This code exists to study how Anthropic's OAuth protocol and Claude Code's HTTP communication pattern work. It is not intended to be used in violation of Anthropic's Terms of Service.

Account risk. There is a real risk of account suspension if Anthropic detects requests are not coming from Claude Code. The mimicry tries hard to make every request look as close as possible to a genuine Claude Code request, but that reduces rather than eliminates the risk. Use at your own risk.

Prerequisites: OpenCode installed, a Claude Pro or Max subscription, Node.js 18+.

# 1. Clone and install git clone https://github.com/marco-jardim/opencode-anthropic-fix.git cd opencode-anthropic-fix npm install # 2. Install the plugin + CLI npm run install:link # 3. Start OpenCode and connect opencode # Press Ctrl+K → Connect Provider → Anthropic → "Claude Pro/Max (multi-account)" # Follow the OAuth prompts to log in

That's it. OpenCode will now use your Claude subscription directly. All model costs show as $0.00.

The original plugin provided basic OAuth support. This fork adds:

• Multi-account support — add up to 10 Claude accounts and rotate between them
• Automatic rate limit handling — when one account hits a limit, the plugin switches to another
• Health scoring — tracks account reliability and prefers healthy accounts
• Standalone CLI — manage accounts without opening OpenCode
• Configurable strategies — sticky, round-robin, or hybrid account selection
• Claude Code signature emulation — full HTTP header, system prompt, beta flag, and metadata mimicry derived from Claude Code's open source code
• OAuth endpoint fingerprint parity — sends Claude Code-style User-Agent on /v1/oauth/token and /v1/oauth/revoke to match current OAuth validation
• Effort-based thinking for Opus 4.6 — maps budgetTokens to effort levels (low/medium/high) and includes effort-2025-11-24
• Upstream-aligned auto betas — advanced-tool-use-2025-11-20 and fast-mode-2026-02-01 auto-included to match Claude Code 2.1.79+ (redact-thinking-2026-02-12 available as opt-in to preserve thinking block visibility)
• 1M context limit override — patches model.limit.context so OpenCode compacts at the right threshold while models.dev catches up
• Runtime config + custom betas — /anthropic set, /anthropic config, and /anthropic betas slash commands for live feature toggling without restarting OpenCode
• Files API integration — upload, list, download, and manage files via /anthropic files with endpoint/content-scoped files-api-2025-04-14 beta injection
• Code execution support — available via explicit custom beta opt-in (code-execution-2025-08-25), not auto-enabled

When Claude Code is installed and authenticated on the same machine, this plugin can reuse its OAuth credentials instead of creating separate tokens. This means:

• Same token: Anthropic sees exactly one client — Claude Code — with no second login event
• Zero detection surface: No duplicate token, no correlation between separate OAuth flows
• Automatic: Credentials are detected and loaded on plugin startup

• Claude Code must be installed (claude command available)
• Claude Code must be authenticated (claude login has been run)
• On macOS: credentials stored in Keychain (automatic)
• On Linux: credentials stored in ~/.claude/.credentials.json

Control the feature via ~/.config/opencode/anthropic-auth.json:

{ "cc_credential_reuse": { "enabled": true, // Enable/disable the feature (default: true) "auto_detect": true, // Auto-load on plugin startup (default: true) "prefer_over_oauth": true, // Use CC credentials first when available (default: true) }, }

If you prefer to use separate OAuth tokens:

# Via config file echo '{"cc_credential_reuse":{"enabled":false}}' > ~/.config/opencode/anthropic-auth.json

Or set via environment variable (disables auto-detection):

export OPENCODE_ANTHROPIC_CC_REUSE_ENABLED=false

"No Claude Code credentials found"

• Claude Code is not installed or not authenticated
• Run claude login first, then retry

Keychain permission prompts

• First read may trigger macOS Allow/Deny dialog
• Click "Allow" to grant access
• If denied, credentials will fall back to file reading

Token expiry

• CC credentials are refreshed by invoking claude -p . --model haiku
• This triggers CC's own token refresh mechanism
• If Claude Code is not in PATH, refresh will fail and account rotation will occur

Add to your opencode.json:

{ "plugins": ["opencode-anthropic-fix@latest"] }

OpenCode will install and load the plugin automatically on next start.

Best for active development. Edits to source files take effect immediately.

npm run install:link

This creates:

• Plugin: ~/.config/opencode/plugin/opencode-anthropic-auth-plugin.js → ./index.mjs
• CLI: ~/.local/bin/opencode-anthropic-auth → ./cli.mjs

Bundles the plugin and CLI into self-contained single files (via esbuild) and copies them. No symlinks, no node_modules needed at the destination.

npm run install:copy

This creates:

• Plugin: ~/.config/opencode/plugin/opencode-anthropic-auth-plugin.js (standalone, ~50KB)
• CLI: ~/.local/bin/opencode-anthropic-auth (standalone, ~35KB)

npm run uninstall

If ~/.local/bin isn't on your PATH, add it:

export PATH="$HOME/.local/bin:$PATH"

1. Open OpenCode
2. Press Ctrl+K → Connect Provider → Anthropic
3. Select "Claude Pro/Max (multi-account)"
4. Open the URL in your browser, authorize, and paste the code back

You can also add accounts directly from the CLI without opening OpenCode:

opencode-anthropic-auth login

Run the auth flow again (via CLI login or OpenCode's Connect Provider). The plugin detects existing accounts and shows a menu:

2 account(s) configured: 1. alice@example.com (active) 2. bob@example.com (a)dd new, (f)resh start, (m)anage, (c)ancel? [a/f/m/c]: 

• Add — log in with another Claude account
• Fresh start — clear all accounts and start over
• Manage — enable/disable/remove accounts inline
• Cancel — keep current setup

This fork is OAuth-first. Use claude.ai login flows (login / reauth) for all accounts.

The CLI lets you manage accounts outside of OpenCode.

opencode-anthropic-auth [command] [args]

# Add a new account via browser OAuth opencode-anthropic-auth login # See account status (includes live usage quotas) opencode-anthropic-auth list # Output: # Anthropic Multi-Account Status # # Account Status Failures Rate Limit # ────────────────────────────────────────────────────────────── # 1 alice@example.com ● active 0 — # 5h █████░░░░░ 45% resets in 2h 30m # 7d █████░░░░░ 45% resets in 4d 16h # Sonnet 7d ░░░░░░░░░░ 0% # # 2 bob@example.com ● ready 0 — # 5h ███░░░░░░░ 31% resets in 30m # 7d ███████░░░ 70% resets in 19h 31m # Sonnet 7d █░░░░░░░░░ 11% resets in 2d 10h # # Strategy: sticky | 2 of 2 enabled # Storage: ~/.config/opencode/anthropic-accounts.json # Switch active account opencode-anthropic-auth switch 2 # Re-authenticate a broken account (opens browser) opencode-anthropic-auth reauth 1 # Quick token refresh without browser opencode-anthropic-auth refresh 1 # Revoke tokens and remove an account opencode-anthropic-auth logout 2 # Revoke all tokens and clear all accounts opencode-anthropic-auth logout --all # View token usage per account opencode-anthropic-auth stats # Reset all usage counters opencode-anthropic-auth reset-stats all # One-liner for shell prompts opencode-anthropic-auth status # anthropic: 2 accounts (2 active), strategy: sticky, next: #1 # Interactive management opencode-anthropic-auth manage

Most commands have short aliases: ln, lo, ra, rf, ls, st, sw, en, dis, rm, strat, cfg, mg.

The plugin registers a built-in /anthropic slash command for account management, feature toggles, and custom beta headers — all without leaving OpenCode.

/anthropic # list accounts (default) /anthropic usage # full account list + quota windows /anthropic switch 2 /anthropic refresh 1 /anthropic logout 2 # revoke tokens and remove account 2 /anthropic logout --all # revoke all tokens and clear all accounts /anthropic stats 

Toggle features live without restarting OpenCode. Changes are persisted to anthropic-auth.json.

/anthropic config # show all current settings /anthropic set emulation on # enable/disable Claude signature emulation /anthropic set compaction minimal # set prompt compaction mode (minimal/off) /anthropic set 1m-context on # enable/disable 1M context limit override /anthropic set idle-refresh on # enable/disable idle account refresh /anthropic set strategy round-robin # change account selection strategy /anthropic set debug on # enable/disable debug logging /anthropic set quiet on # suppress non-error toasts 

Add or remove beta flags that get included in every anthropic-beta header. Persisted across sessions.

/anthropic betas # show active betas (auto + custom) with presets /anthropic betas add <beta-name> # add a custom beta /anthropic betas remove <beta-name> # remove a custom beta /anthropic betas add 1m # shortcut => context-1m-2025-08-07 

Available preset betas (shown by /anthropic betas):

Example workflow:

/anthropic betas add prompt-caching-scope-2026-01-05 /anthropic betas add context-management-2025-06-27 /anthropic betas # verify 

Upload, list, download, and delete files via the Anthropic Files API. The files-api-2025-04-14 beta is injected only when needed (/v1/files requests or Messages payloads with file_id).

/anthropic files # list files across ALL accounts /anthropic files upload ./report.pdf # upload a file /anthropic files get file_abc123 # get file metadata /anthropic files download file_abc123 # download to current directory /anthropic files download file_abc123 ./out.pdf # download to specific path /anthropic files delete file_abc123 # delete a file 

Supported formats: PDF, DOCX, TXT, CSV, Excel, Markdown, images (max 350 MB per file). Uploaded files can be referenced by file_id in Messages API requests.

Multi-account behavior: Files on Anthropic are per-account. With multiple accounts:

• /anthropic files list (no --account) queries all enabled accounts, labeling each file with its owner email
• Use --account <email|index> to target a specific account for any action:

  /anthropic files list --account alice@example.com /anthropic files upload ./data.csv --account 2 

• Auto-pinning: When you upload or list files, the plugin remembers which account owns each file_id. If a subsequent Messages API request references that file_id, the plugin automatically routes it to the correct account — even with round-robin or hybrid strategies.

Login and reauth are two-step flows in slash mode:

/anthropic login # opens URL instructions in chat /anthropic login complete <code#state> /anthropic reauth 1 # opens URL instructions in chat /anthropic reauth complete <code#state> 

Pending slash OAuth flows expire after 10 minutes. If completion fails with an expiration message, run the start command again.

• Destructive commands (remove, logout) run with non-interactive --force behavior in slash mode.
• Interactive manage is terminal-only; use granular slash commands instead.

Control how the plugin picks which account to use for each request.

# Via CLI opencode-anthropic-auth strategy round-robin # Via environment variable (overrides config file) export OPENCODE_ANTHROPIC_STRATEGY=hybrid # Via config file # Edit ~/.config/opencode/anthropic-auth.json

Some Anthropic API features maintain server-side per-account state that breaks when requests alternate between accounts:

Recommendation for round-robin users: If you need prompt caching or code execution, pin each OpenCode session to a single account:

# Terminal 1 — uses account 1 OPENCODE_ANTHROPIC_INITIAL_ACCOUNT=1 opencode # Terminal 2 — uses account 2 OPENCODE_ANTHROPIC_INITIAL_ACCOUNT=2 opencode

This automatically overrides the strategy to sticky for that session, re-enabling strategy-sensitive auto betas (for example prompt-caching-scope-2026-01-05). Other sessions are unaffected.

Configuration is stored at ~/.config/opencode/anthropic-auth.json. All settings are optional — defaults work well for most users.

{ // Account selection strategy: "sticky" | "round-robin" | "hybrid" "account_selection_strategy": "sticky", // Seconds before consecutive failure count resets (60-7200) "failure_ttl_seconds": 3600, // Enable debug logging "debug": false, // Claude Code signature emulation behavior "signature_emulation": { // Enable Claude-style attribution/stainless headers and betas "enabled": true, // Resolve latest @anthropic-ai/claude-code version once on plugin startup "fetch_claude_code_version_on_startup": true, // Compact long injected system instructions to reduce token usage. // In "minimal" mode, repeated/contained blocks are deduplicated and title-generator // requests are replaced with a compact dedicated prompt. // "minimal" | "off" "prompt_compaction": "minimal", }, // Context limit override for 1M-window models. // Prevents OpenCode from compacting too early when models.dev hasn't been // updated yet (e.g. claude-opus-4-6 and any *-1m model variants). // Only applied for OAuth (Max Plan) sessions — API key users use the // context-1m-2025-08-07 beta header instead. "override_model_limits": { // Enable/disable the override (default: off — enable if you need 1M context) "enabled": false, // Context window to inject (tokens). Default: 1_000_000. "context": 1000000, // Max output tokens to inject. 0 = leave the model's default unchanged. "output": 0, }, // Health score tuning (0-100 scale) "health_score": { "initial": 70, "success_reward": 1, "rate_limit_penalty": -10, "failure_penalty": -20, "recovery_rate_per_hour": 2, "min_usable": 50, "max_score": 100, }, // Client-side rate limiting (token bucket) "token_bucket": { "max_tokens": 50, "regeneration_rate_per_minute": 6, "initial_tokens": 50, }, // Custom beta headers (added to every request via /anthropic betas add) "custom_betas": [], // Toast notification settings "toasts": { // Suppress non-error toasts (account status, switching) "quiet": false, // Minimum seconds between account-switch toasts (0-300) "debounce_seconds": 30, }, }

• The expected auth mode is OAuth (claude login / browser flow), not direct ANTHROPIC_API_KEY usage.
• In OAuth mode, the plugin always includes oauth-2025-04-20 in anthropic-beta.
• This applies to all models, including Haiku.

When you make a request through OpenCode:

1. The plugin selects an account based on your strategy
2. It refreshes the OAuth token if expired
3. It transforms the request (adds OAuth headers, oauth-2025-04-20, signature headers, beta flags, tool prefixes)
4. If the response is account-specific (429/401, plus 400/403 billing/quota/permission errors), it marks that account and immediately tries the next account
5. If the response is service-wide (500/503/529), it returns the error directly (switching accounts would not help)
6. It tries each available account at most once per request
7. Successful responses have tool name prefixes stripped from the stream

The plugin also:

• Zeros out model costs (your subscription covers usage)
• Emulates Claude-style request headers and beta flags by default
• Sanitizes "OpenCode" references to "Claude Code" in system prompts (required by Anthropic's API)
• In prompt_compaction="minimal", deduplicates repeated/contained system blocks and uses a compact dedicated prompt for internal title-generation requests
• Adds ?beta=true to /v1/messages and /v1/messages/count_tokens requests

When signature emulation is disabled (signature_emulation.enabled=false), the plugin falls back to legacy behavior including the Claude Code system prompt prefix.

Account credentials are stored with restrictive file permissions (owner read/write only) and are excluded from git via an auto-generated .gitignore.

Make sure the plugin is installed in ~/.config/opencode/plugin/. Restart OpenCode after installing.

Your OAuth token may have expired. Try a quick refresh first, or re-authenticate with fresh browser login:

# Quick token refresh (no browser needed) opencode-anthropic-auth refresh 1 # Full re-authentication (opens browser) opencode-anthropic-auth reauth 1

Or re-run the auth flow from OpenCode: Ctrl+K → Connect Provider → Anthropic.

Enable system prompt debug logging:

export OPENCODE_ANTHROPIC_DEBUG_SYSTEM_PROMPT=1 opencode

When enabled, the plugin prints the transformed system block (after sanitization/compaction) to stderr with prefix:

[opencode-anthropic-auth][system-debug] transformed system: 

Note: internal title-generator requests are intentionally skipped by this debug log to avoid noisy high-volume output.

When all accounts are exhausted for an account-specific error, the plugin returns immediately instead of sleeping in-process. Try:

# Check status opencode-anthropic-auth list # Reset tracking if stuck opencode-anthropic-auth reset all

You may see Claude: <email> toasts even when your selected model is not Anthropic (e.g., OpenAI Codex, Gemini). This is expected behavior — OpenCode uses Claude Haiku as a background "small model" for internal tasks like generating session titles, regardless of which model you selected. These background API calls go through the Anthropic provider, which triggers the plugin's fetch interceptor and its account-usage toast.

To suppress non-error toasts, set quiet mode in your config:

// ~/.config/opencode/anthropic-auth.json { "toasts": { "quiet": true }, }

Or via environment variable:

export OPENCODE_ANTHROPIC_QUIET=1

Error toasts (e.g., "Disabled Account 1 (token refresh failed)") are never suppressed.

Make sure ~/.local/bin is on your PATH:

export PATH="$HOME/.local/bin:$PATH"

Same as upstream. See anomalyco/opencode-anthropic-auth.

Maintained at marco-jardim/opencode-anthropic-fix.