New API

New API is an AGPL-3.0 LLM gateway by QuantumNous that began as a feature-expanded fork of songquanpeng's one-api and has surpassed the upstream in development velocity and feature scope. It keeps the multi-provider OpenAI-compatible proxy, token wallet, and quota system from one-api while adding OIDC SSO, a redesigned admin UI, expanded model support, and modernized infrastructure.

What it does

New API fronts 40+ providers — OpenAI, Anthropic, Gemini, Azure, AWS Bedrock, DeepSeek, Moonshot, Zhipu, Alibaba, Baidu, iFlytek, and OpenAI-compatible endpoints — behind one OpenAI-compatible API. Channel management lets operators register multiple upstream keys per provider with weighted load balancing and failover. Users hold token-denominated wallets; requests deduct tokens at configurable per-model rates, and requests are rejected at the API layer when the wallet runs out.

The admin UI provides per-user usage dashboards, per-channel analytics, and invitation flows. OIDC SSO plugs into Keycloak, Zitadel, Authentik, and other generic providers for admin-UI authentication.

Where it fits

Deployments that want a self-hostable OpenAI-compatible service with user accounts, wallets, and billing built in. The tool is widely used by small API-reseller businesses as well as internal corporate LLM platforms that want per-employee usage accounting.

Licensing

AGPL-3.0. Internal self-hosted use without redistribution is unencumbered. Exposing a modified version as a publicly accessible network service triggers the AGPL network clause and requires source publication.

Deployment

Single Docker container plus MySQL, Postgres, or SQLite, and optional Redis. Runs on Docker Swarm and Kubernetes. Documentation is bilingual Chinese and English with reasonable English coverage of install and configuration.

Limitations

• AGPL-3.0 makes redistribution as part of a modified commercial product legally complex; teams planning that path should consult counsel.
• No MCP gateway — the tool is a pure LLM proxy.
• Community is centered on Chinese developers; English GitHub issue turnaround can be slower than Western-origin projects.
• Token-wallet model is richer than most enterprises need — internal teams often ignore the billing features and use flat per-user quotas.