elizaos--eliza/plugins/plugin-anthropic-proxy

@elizaos/plugin-anthropic-proxy

Routes Anthropic API traffic from your eliza agent through a Claude Max / Pro subscription instead of paying per-token Extra Usage rates.

The plugin applies a 7-layer bidirectional transformation pipeline so requests look like they originate from the official Claude Code CLI:

1. Billing header injection (x-anthropic-billing-header text block carrying the CC version plus a 3-character SHA256 fingerprint computed per request)
2. String trigger sanitization
3. Tool name fingerprint bypass (PascalCase CC convention rename)
4. System prompt template bypass (strip + paraphrase)
5. Tool description stripping (reduce schema fingerprint)
6. Schema property name renaming
7. Full bidirectional reverse mapping on SSE + JSON responses

Plus assistant-prefill stripping and thinking-block stripping for replay/session bugs.

The default fingerprint dictionaries target the elizaOS tool surface (@elizaos/native-reasoning). For non-eliza agents, supply your own dictionaries via config.json (see config.json.example).

Custom fingerprint dictionaries

The defaults make this plugin a one-line drop-in for any eliza agent. If you're running a non-eliza agent (LangChain, LlamaIndex, your own runtime, etc.) the eliza tool-name dictionary won't match your tool surface and you'll want to supply your own.

Drop a config.json next to your eliza root with the shape shown in config.json.example. The plugin merges it over the defaults at startup. Any of the four dictionaries (replacements, toolRenames, propRenames, reverseMap) can be overridden independently — the rest fall back to the eliza defaults.

You own the subscription

This plugin does not route your traffic through any service operated by anyone but you. It needs your Claude Code OAuth token (from your own subscription on your own machine). You are responsible for whether your usage complies with Anthropic's terms.

Setup

# 1. Install Claude Code CLI and log in once on this machine.
claude auth login

# 2. Add the plugin to your agent's plugin list (your character file or
#    plugin loader). It will:
#    - Start an in-process proxy on http://127.0.0.1:18801
#    - Set ANTHROPIC_BASE_URL to that proxy URL (unless you've set it
#      explicitly to something else)

Modes

Pick via CLAUDE_MAX_PROXY_MODE:

Mode	What it does
inline	(default) Plugin starts an http proxy in this agent's process
shared	Plugin connects to an existing upstream proxy URL (one host, many agents)
off	Plugin loads but doesn't start anything (passthrough; you set ANTHROPIC_BASE_URL yourself)

In inline mode each agent gets its own proxy server. In shared mode you run the proxy once on the host (or via this same plugin in a different agent) and point all your agents at the same CLAUDE_MAX_PROXY_UPSTREAM. Useful when you have many agents on one box and only one Claude subscription.

Environment variables

Variable	Default	Notes
CLAUDE_MAX_PROXY_MODE	inline	inline / shared / off
CLAUDE_MAX_PROXY_PORT	18801	inline mode listen port
CLAUDE_MAX_PROXY_UPSTREAM	(none)	shared mode upstream base URL, e.g. http://172.18.0.1:18801
CLAUDE_MAX_PROXY_BIND_HOST	127.0.0.1	inline mode bind address
CLAUDE_MAX_PROXY_AUTH_TOKEN	(none)	required when CLAUDE_MAX_PROXY_BIND_HOST is not a loopback address; checked via Authorization: Bearer <token> or x-claude-max-proxy-token header
CLAUDE_MAX_PROXY_VERBOSE	false	extra request logging
CLAUDE_MAX_PROXY_CONFIG_PATH	(none)	explicit path to a config.json fingerprint override file; takes precedence over a config.json found next to the agent root
CLAUDE_MAX_CREDENTIALS_PATH	(auto)	path to .credentials.json; defaults to ~/.claude/.credentials.json
CLAUDE_CODE_OAUTH_TOKEN	(none)	direct OAuth bearer token; takes precedence over the file
ANTHROPIC_BASE_URL	(auto-set by plugin)	leave unset and the plugin picks. Set to auto to opt back in if you ever set it. Set to anything else and the plugin will leave it alone.

Diagnostics

• HTTP route: GET /api/anthropic-proxy/status returns the current mode, URL, listening state, request count, token expiry, and (in shared mode) upstream reachability.
• Action: PROXY_STATUS returns the same info to a chat surface.
• Local proxy health: GET http://127.0.0.1:18801/health (replace port to match config).

Token refresh

If you hit a 401 (token expired) run:

claude auth login

The plugin re-reads the credentials file on every request, so a fresh login is picked up immediately — no need to restart the agent.

Failure modes (intentional)

• Missing credentials. Plugin logs a warning, degrades to off mode, agent keeps running. It does not crash.
• Inline port collision. Plugin logs the bind error, degrades to off mode.
• Shared upstream unreachable at startup. Plugin still boots in shared mode; the unreachable upstream is reported via /api/anthropic-proxy/status.

Plugin shape

• services: [AnthropicProxyService] — Service that owns the http server lifecycle (start/stop)
• actions: [proxyStatusAction] — PROXY_STATUS action for in-chat diagnostics
• routes: anthropicProxyRoutes — GET /api/anthropic-proxy/status for external tools
• init() — sets ANTHROPIC_BASE_URL if you haven't already

Files

plugins/plugin-anthropic-proxy/
├── index.ts                           # Plugin export + init
├── index.node.ts                      # Node entry
├── index.browser.ts                   # Browser-unavailable entry
├── auto-enable.ts                      # shouldEnable() opt-in check
├── config.json.example                # Custom fingerprint dictionary shape
├── build.ts                           # Bun build script
├── package.json
├── tsconfig.json / tsconfig.build.json
├── vitest.config.ts
├── bunfig.toml
├── src/
│   ├── proxy/
│   │   ├── constants.ts               # Algorithm constants + DEFAULT_* dict re-exports
│   │   ├── eliza-fingerprint.ts       # ELIZA_* default fingerprint dictionaries
│   │   ├── billing-fingerprint.ts     # Layer 1: SHA256 fingerprint
│   │   ├── sanitize.ts                # Layer 2: string sanitize
│   │   ├── tool-rename.ts             # Layer 3/6: quoted name renames
│   │   ├── system-prompt.ts           # Layer 4: system strip + paraphrase
│   │   ├── cc-tool-injection.ts       # Layer 5: description strip + synthetic CC tools
│   │   ├── sse-rewrite.ts             # Tail-buffer SSE reverse map
│   │   ├── stainless-headers.ts       # CC-emulating SDK headers
│   │   ├── process-body.ts            # Forward request pipeline
│   │   ├── reverse-map.ts             # Response/SSE reverse pipeline
│   │   └── server.ts                  # http.createServer lifecycle
│   ├── services/
│   │   └── proxy-service.ts           # AnthropicProxyService extends Service
│   ├── actions/
│   │   └── proxy-status.action.ts     # PROXY_STATUS action
│   ├── routes/
│   │   └── status-route.ts            # /api/anthropic-proxy/status
│   └── utils/
│       └── credentials-loader.ts      # ~/.claude/.credentials.json + JWT exp
└── __tests__/
    ├── proxy.test.ts
    ├── auto-enable.test.ts
    ├── eliza-fingerprint.test.ts
    ├── manifest-engine.integration.test.ts
    ├── process-body.edge.test.ts
    ├── proxy-server.routing.test.ts
    └── sse-rewrite.test.ts