- Containers:
HOST/AXIORANK_GATEWAY_HOSTenv vars now override the bind host (the CLI default stays loopback). The Docker image setsHOST=0.0.0.0so a published port is actually reachable. - The Dockerfile builds from source in a builder stage, so
docker buildworks straight from a clone, and a tagged release publishesghcr.io/axiorank/gateway(multi-arch) from CI. - Runnable examples under
examples/and one-command deploys (docker compose, Render, Fly.io).
- 22 new provider presets, bringing the total to 34. Hosted: cerebras, sambanova, deepinfra, novita, hyperbolic, nebius, moonshot (Kimi), zhipu (GLM via Z.ai), qwen (Alibaba DashScope), nvidia (NIM), huggingface (Inference Providers router), github (GitHub Models), cohere, baseten, featherless, siliconflow, scaleway, and minimax. Local, no API key: llamacpp, sglang, koboldcpp, and jan. Every base URL is verified against the provider's official docs and smoke tested.
- Cost-routing price snapshot refreshed from the AxioRank model catalog.
initnow prints the docs and source links in its next steps, and the offlinedemoclosing panel carries the repository URL.
First release. An npx-runnable, OpenAI-compatible AI gateway with:
- Guardrails on by default (block, redact, observe, off) on both the prompt and the completion, powered by the AxioRank detection engine, running fully offline on the hot path.
- A signed, hash-chained Gateway Receipt on every response (including blocked ones), verifiable offline with the public JWKS.
- Routing across providers with failover, per-route retries, and per-attempt timeouts. Strategies: failover, cost, round_robin.
- First-class upstreams openai, azure, openrouter, and custom, plus presets for groq, together, fireworks, mistral, deepseek, xai, perplexity, anthropic, gemini, ollama, vllm, and lmstudio.
- Exact-match in-memory response cache, opt-in per route.
- Optional AxioRank Cloud sync (
pull/push) and receipt telemetry whenAXIORANK_KEYis set. - CLI: start, init, keys, verify, pull, push, and an offline
demo. - Zero runtime dependencies.