Claude Proxy

Multi-provider API proxy with an Anthropic-compatible endpoint. Drop-in replacement for api.anthropic.com -- routes requests through OpenCode with automatic fallback across 14 providers. When no OpenCode API key is configured, Puter is used as primary (free Claude models).

Quick Start

cp .env.example .env   # add your API keys
npm install
npm start              # starts on http://0.0.0.0:4000

Point Claude Code at the proxy (~/.claude/settings.json):

{
  "env": {
    "ANTHROPIC_API_KEY": "YOUR_API_KEY",
    "ANTHROPIC_BASE_URL": "http://localhost:4000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Architecture

Client
  |  POST /v1/messages (Anthropic format)
  v
Express server
  |-- routeFor()           -- map model name to backend model ID
  |-- getModelConfig()     -- pick endpoint + protocol
  |-- sanitizeMessages()   -- repair JSON, coerce roles, strip surrogates
  |-- buildChain()         -- build 14-provider fallback chain
  |-- tryChain()           -- parallel (opus=3, sonnet=2) then sequential
  |     |-- cooldown-aware skip
  |     |-- rate-limit header parsing + bucket state tracking
  |     |-- circuit breaker (3 consecutive failures -> 5min cooldown)
  |     |-- jittered backoff (1s, 2s, 4s, 8s)
  |     |-- content-type validation (detect error-page HTML)
  |     |-- per-provider TTFB timeouts (8s-25s)
  |     |-- fetch with proxy agent
  |     v
  |-- format conversion     -- openaiToAnthropic / anthropicToOpenAI
  |     |-- streaming think scrubber (<think>/<reasoning> at SSE boundaries)
  |     |-- mid-stream silent retry
  |     |-- Ollama tool-call slot fix
  |-- logRequest()          -- persist to SQLite (degradation-safe)
  v
Response (Anthropic format)

Providers

#	Provider	Key Required	Tier	Notes
1	OpenCode	Yes	Primary	Skipped if OPENCODE_API_KEY is unset
2	Puter	Yes (token)	Free	Primary when no OpenCode key; uses native Claude IDs (claude-sonnet-4-6, claude-opus-4-8, claude-haiku-4-5); 200k context
3	Ollama	No	Local	Self-hosted, no rate limits
4	NVIDIA NIM	No	Free	Rate-limited, free tier
5	Kimi (Moonshot)	Yes	Fallback	256k context; models mapped as kimi-k2.*
6	OpenRouter	Yes	Fallback	Broad model selection
7	Cerebras	Yes	Fallback	Fast inference
8	GitHub Models	Yes	Fallback	Azure-backed
9	Mistral	Yes	Fallback	Le Chat Mistral
10	HuggingFace	Yes	Fallback	Serverless Inference API
11	Novita	Yes	Fallback	Pay-as-you-go
12	OpenAI	Yes	Fallback	Quota-limited
13	Gemini	Yes	Fallback	Quota-limited
14	Groq	Yes	Fallback	Fast but rate-limited; tried last

Configuration

Variable	Default	Description
SECRET_KEY	--	API key clients send as x-api-key
OPUS_MAP_MODEL	deepseek-v4-flash-free	Model for opus requests
SONNET_MAP_MODEL	mimo-v2.5-free	Model for sonnet requests
HAIKU_MAP_MODEL	north-mini-code-free	Model for haiku requests
PRIMARY_PROVIDER	--	Force a specific provider first (e.g. puter)
PUTER_AUTH_TOKEN	--	JWT token from puter.com/dashboard for Puter API
KIMI_API_KEY / MOONSHOT_API_KEY	--	API key for Kimi (Moonshot)
HOST	0.0.0.0	Bind address
PORT	4000	Listen port
OPENCODE_BASE_URL	https://opencode.ai	Upstream OpenCode API base URL
OPENCODE_PROXY	--	HTTP proxy URL for upstream requests

API Endpoints

Method	Path	Description
POST	/v1/messages	Anthropic Messages API (main)
POST	/anthropic/v1/messages	Alias for /v1/messages
GET	/v1/models	List configured models
GET	/v1/providers	List providers with health status
GET	/health	Health check (Docker-compatible)
GET	/dash	Live web dashboard
GET	/dash/data	Dashboard stats (JSON)

Tech Stack

Runtime: Node.js 20+ | Language: TypeScript (via tsx) | Framework: Express 5 | DB: SQLite (better-sqlite3) | TUI: blessed | Dashboard: Server-side HTML with ANIMO dark theme (#0f131c / #caf979)

License

MIT