GPT-Load
A self-hosted AI gateway pools API keys and subscription accounts, with automatic failover and support for OpenAI, Anthropic and Gemini protocols.

GPT-Load is a self-hosted AI gateway for developers connecting applications to multiple AI providers and accounts. It centralizes credentials and routing behind one endpoint, so each application doesn't need its own failure handling.
- APIs and subscriptions: Pool API keys alongside Codex, Claude, Antigravity and Grok subscription accounts, with shared scheduling and credential health checks.
- Client compatibility: Connect clients using OpenAI Chat Completions, OpenAI Responses, Anthropic Messages or Gemini protocols through native or converted routes, depending on the channel.
- Traffic and failover: Distribute requests with weighted routing and session affinity, retry failures, and temporarily exclude overloaded or invalid credentials.
- Logs and usage: Inspect routes, request logs and credential health in the built-in management UI, with usage summaries and per-model cost estimates.
- Application access: Control which provider groups and protocols each application can use, with rate and cost limits on its access key.
GPT-Load is open source under the MIT License. It runs through Docker or native binaries on Linux, macOS and Windows, and supports SQLite, MySQL and PostgreSQL. Credentials are encrypted locally, while AI requests go to your chosen upstream services; self-hosting the gateway doesn't make those services local or offline.
Protocol and operation support varies by channel. Usage and cost figures are estimates rather than billing records. Each deployment supports a single application instance, without shared state or horizontal scaling.