Description
You end up holding a pile of keys: one for OpenAI, one for Claude, one for Gemini, one for DeepSeek. Who on the team is using which, how much each cost this month, which channel is rate-limited right now — all of it lives in someone’s head. And giving a colleague access means sending them the raw key, which you can never take back.
New API is a gateway that sits in front of all of those upstreams: configure every channel in one place and expose a single OpenAI-compatible endpoint, so your applications only ever know one address and one token. It also converts between formats — call Claude using the OpenAI format, or come in through native Claude Messages or Google Gemini and let the gateway translate, with no client-side changes.
Tokens are governed in detail: expiry, quota, allowed IP ranges, and which models they may reach. Requests spread across channels by weighted random selection with automatic retry on failure and per-user rate limiting. Billing is accounted per request, by usage and by cache hit, with cache-billing statistics for OpenAI, Azure, DeepSeek, Claude, Qwen and others.
It is a fork of One API and remains fully compatible with the original database, so an existing deployment migrates straight across. Written in Go as a single binary, one line to run under Docker, SQLite by default and MySQL or PostgreSQL when volume demands it. AGPL-3.0, 47k stars.
One entry point: OpenAI, Claude, Gemini, DeepSeek and others become channels behind a single compatible endpoint that your applications connect to.
Format conversion: OpenAI-compatible ⇄ Claude Messages both ways, OpenAI-compatible → Google Gemini, Gemini → OpenAI-compatible (text only, function calling not yet supported), plus thinking-to-content handling.
Reasoning effort in the model name:
Broad interface coverage: Chat Completions, Responses, Realtime (Azure included), Claude Messages, native Gemini, images, audio, video (Sora), embeddings and rerank (Cohere, Jina), plus Midjourney-Proxy, Suno API and Dify ChatFlow.
Tokens and permissions: per-token expiry, quota, allowed IP ranges and model allowlists, combined with user groups, channel groups and per-group rate multipliers for multi-tenant setups.
Smart routing: weighted random channel selection, automatic retry on failure, and user-level model rate limiting.
Billing and analytics: per-request, usage-based and cache-hit accounting with a visual console for quota detail, and EPay or Stripe for internal top-ups and quota allocation.
Many sign-in methods: email, GitHub, Discord, Telegram and LinuxDO authorisation logins, plus OIDC single sign-on.
Interface: a redesigned UI with a visual dashboard, available in Simplified Chinese, Traditional Chinese, English, French and Japanese.
Deployment: the
A compliance note: the project states in its own README that upstream keys and model services must be obtained lawfully, and that anyone offering generative AI services to the public must first meet the filing, content-safety, real-name verification and log-retention obligations of their jurisdiction. Worth reading properly before putting a resale gateway on the public internet.
New API is a gateway that sits in front of all of those upstreams: configure every channel in one place and expose a single OpenAI-compatible endpoint, so your applications only ever know one address and one token. It also converts between formats — call Claude using the OpenAI format, or come in through native Claude Messages or Google Gemini and let the gateway translate, with no client-side changes.
Tokens are governed in detail: expiry, quota, allowed IP ranges, and which models they may reach. Requests spread across channels by weighted random selection with automatic retry on failure and per-user rate limiting. Billing is accounted per request, by usage and by cache hit, with cache-billing statistics for OpenAI, Azure, DeepSeek, Claude, Qwen and others.
It is a fork of One API and remains fully compatible with the original database, so an existing deployment migrates straight across. Written in Go as a single binary, one line to run under Docker, SQLite by default and MySQL or PostgreSQL when volume demands it. AGPL-3.0, 47k stars.
Features
One entry point: OpenAI, Claude, Gemini, DeepSeek and others become channels behind a single compatible endpoint that your applications connect to.
Format conversion: OpenAI-compatible ⇄ Claude Messages both ways, OpenAI-compatible → Google Gemini, Gemini → OpenAI-compatible (text only, function calling not yet supported), plus thinking-to-content handling.
Reasoning effort in the model name:
o3-mini-high or gpt-5-low for OpenAI, claude-3-7-sonnet-20250219-thinking for Claude, -thinking / -nothinking and thinking budgets for Gemini, or simply append -low, -medium or -high.Broad interface coverage: Chat Completions, Responses, Realtime (Azure included), Claude Messages, native Gemini, images, audio, video (Sora), embeddings and rerank (Cohere, Jina), plus Midjourney-Proxy, Suno API and Dify ChatFlow.
Tokens and permissions: per-token expiry, quota, allowed IP ranges and model allowlists, combined with user groups, channel groups and per-group rate multipliers for multi-tenant setups.
Smart routing: weighted random channel selection, automatic retry on failure, and user-level model rate limiting.
Billing and analytics: per-request, usage-based and cache-hit accounting with a visual console for quota detail, and EPay or Stripe for internal top-ups and quota allocation.
Many sign-in methods: email, GitHub, Discord, Telegram and LinuxDO authorisation logins, plus OIDC single sign-on.
Interface: a redesigned UI with a visual dashboard, available in Simplified Chinese, Traditional Chinese, English, French and Japanese.
Deployment: the
calciumion/new-api Docker image or Docker Compose; SQLite locally, MySQL 5.7.8+ or PostgreSQL 9.6+ remotely; 64-bit only (amd64/arm64). GitHub Releases also ship single-file binaries for Windows, Linux and macOS.A compliance note: the project states in its own README that upstream keys and model services must be obtained lawfully, and that anyone offering generative AI services to the public must first meet the filing, content-safety, real-name verification and log-retention obligations of their jurisdiction. Worth reading properly before putting a resale gateway on the public internet.
