// the take
gpt-load is a self-hosted Go gateway that sits in front of OpenAI, Anthropic, Gemini and a long list of other providers, letting you pool API keys and subscription accounts (Codex, Claude, Antigravity, Grok) behind one URL with scheduling, failover and usage logging. It's aimed at teams or power users running multiple AI accounts/keys who want load balancing and visibility instead of hardcoding one key per app.
Native multi-protocol support is the real feature: clients keep using their normal OpenAI/Anthropic/Gemini request shapes while gpt-load handles routing underneath, so you're not stuck behind a lossy translation layer. Scheduling is genuinely built out — weighted retries, cooldown, blacklisting and session affinity across credentials, not just round-robin. It treats subscription OAuth accounts (Codex/Claude/Antigravity/Grok) as first-class citizens alongside raw API keys in the same credential pool, which matters now that a lot of usage is subscription-based rather than metered API. Storage is pluggable (SQLite out of the box, MySQL/Postgres via DSN for scale) and the test suite is large and specific (golden route tests, per-module channel tests), suggesting the routing logic is actually exercised rather than just hand-waved.
There's no encryption key rotation in this version — lose or need to rotate encryption.key and every stored credential becomes permanently undecryptable, which is a nasty operational trap for anything other than a toy deployment. 2.0 is a full rewrite with zero migration path from 1.x; anyone currently running 1.x has to stand up a parallel deployment and cut over manually rather than upgrade in place. The subscription OAuth flows use fixed callback ports (1455/54545/51121), which caps you at one default instance per host and complicates anything beyond single-box deployment. It's also a single point of failure for all AI traffic across every app routed through it — the README covers swapping to an external DB but says nothing about clustering or HA, so a crash takes down every provider at once, not just one.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →