![\"Aurora](\"https://raw.githubusercontent.com/aurorallm/aurora/main/docs-assets/assets/aurora-logo-animated.svg\"){kind=logo}
\n\n \n
\n ![\"License\"](\"https://img.shields.io/github/license/aurorallm/aurora\")
![\"npm\"](\"https://img.shields.io/npm/v/iaurora\")
![\"npm](\"https://img.shields.io/npm/dw/iaurora\")
![\"GitHub](\"https://img.shields.io/github/stars/aurorallm/aurora\")
\n
![\"Docker](\"https://img.shields.io/docker/pulls/aurorahq/aurora\")
![\"Artifact](\"https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/aurora-gateway\"){kind=icon}
One API for every AI provider. Self-hosted. No vendor lock-in.
\n\n14 provider types • OpenAI & Anthropic compatible • Go • Apache 2.0 • Built for raw speed
\n\n\n![\"Aurora](\"https://raw.githubusercontent.com/aurorallm/aurora/main/docs-assets/assets/dashboard-overview.png\")
\n\n\n---\n\n## What Aurora Does\n\nAurora sits between your app and LLM providers. Your app sends requests using the standard OpenAI or Anthropic SDK — Aurora routes them to whichever provider you've configured. One format handles everything — you dont need to worry about provider-specific formats.\n\n\n```python\n# Before: hardcoded provider\nclient = OpenAI(base_url=\"https://api.openai.com/v1\", api_key=\"sk-...\")\n\n# After: Aurora Gateway\nclient = OpenAI(base_url=\"http://localhost:8080/v1\", api_key=\"your-aurora-key\")\n```\n\nNo SDK changes. No format changes. Just swap the `base_url`.\n\n---\n\n## Features\n\n### Routing & Providers\n\n- **14 provider types** — OpenAI, Anthropic, Gemini, Groq, DeepSeek, OpenRouter, xAI, Z.ai, MiniMax, Azure OpenAI, Oracle, Ollama, vLLM, Jina\n- **Auto-discovery** — set an API key as an env var, restart, provider + all its models appear automatically\n- **Provider pools** — group multiple keys/endpoints, load-balance with round-robin or weighted distribution, health-aware failover\n- **Model aliases** — rename/remap any model to a custom identifier across the entire gateway\n- **Model overrides** — enable or disable specific models per user path, persisted via dashboard or `user_pricing.yaml`\n- **Fallback** — automatic failover on 5xx/429, or manual rules (from config or external JSON) mapping failed provider+model to backups\n- **Resilience** — exponential backoff with jitter, circuit breaker per provider (closed → open → half-open), per-provider override of global retry/circuit-breaker settings\n- **Multiple instances** — run `OPENAI_EAST_API_KEY` and `OPENAI_WEST_API_KEY` as separate providers\n- **Custom base URLs** — override any provider's endpoint (corporate proxies, regional endpoints)\n- **Passthrough** — `/p/{provider}/*` for full upstream API access (not just chat completions); filter which provider types get passthrough routes\n- **Config-driven workflows** — per-request routing, caching, guardrail, audit, usage, budget, and fallback behavior controlled by persisted workflow documents\n\n### API Surface\n\n- **OpenAI-compatible** — `/v1/chat/completions`, `/v1/embeddings`, `/v1/rerank`, `/v1/models`, `/v1/files`, `/v1/batches`\n- **Responses API** — `/v1/responses` with full CRUD, cancel, input items, compact\n- **Anthropic-compatible** — `/v1/messages`, `/v1/messages/count_tokens` (native Anthropic wire format); optional dedicated ingress at `/v1/messages`\n- **Streaming** — SSE streaming for all endpoints, preserved end-to-end\n- **Keep-only-aliases mode** — hide raw provider models from `/v1/models` and expose only aliased names\n- **Configured provider models mode** — `fallback` (add listed models to auto-discovered) or `allowlist` (only serve explicitly listed models)\n\n### Caching\n\n- **Exact cache** — SHA-256 hash match on request, Redis-backed, async writes\n- **Semantic cache** — vector similarity with configurable threshold, supports Qdrant, pgvector, Pinecone, Weaviate\n- **Prompt cache** — forwards `cache_control` to Anthropic/OpenAI/Gemini native prompt caching; configurable modes (`auto`, `manual`, `off`), component toggles, and minimum token threshold\n- **Model registry cache** — local filesystem + Redis, offline-safe; supports vendored JSON snapshots with per-field user pricing overrides\n\n### Security & Guardrails\n\n- **Master key** — top-level gateway auth\n- **Managed API keys** — scoped, rate-limited, per-key model authorization, usage stats\n- **Rate limiting** — per-key rate limiting backed by in-memory or Redis\n- **PII redaction** — email, phone, SSN, credit card detection and masking\n- **Prompt injection blocking** — detects and blocks injection attempts\n- **System prompt protection** — inject, override, or decorate system prompts\n- **Regex blocking** — custom pattern matching with block or sanitize actions\n- **Length limits** — character/token count enforcement on requests\n- **LLM-based altering** — guardrail that rewrites message content via an auxiliary LLM call (anonymization, custom prompts)\n- **Guardrail direction & ordering** — run before provider dispatch (`input`), after response (`output`), or both; same-order guardrails run in parallel\n- **Batch guardrails** — apply configured guardrails to inline items in `/v1/batches` requests\n\n### Observability\n\n- **Audit logging** — full request/response capture, buffered writes, configurable retention (body/header logging, buffer size, flush interval), live SSE stream\n- **Usage analytics** — per-model token counting, cost tracking, daily aggregation by model/user-path, pricing recalculation action\n- **Prometheus metrics** — `aurora_requests_total`, `aurora_request_duration_seconds`, `aurora_requests_in_flight`, plus gateway phase timing\n- **Admin dashboard** — React SPA built into the Go binary: providers, pools, models, aliases, guardrails, cache, usage, audit, auth keys, workflows, console, playground\n- **pprof endpoints** — Go runtime profiling at `/debug/pprof/*` (heap, goroutine, mutex, block, threadcreate)\n- **Structured logging** — configurable format (JSON/text), level (debug/info/warn/error), source info, service metadata\n\n### Cost Control\n\n- **Token saver** — policy-driven output compression (profiles: concise, caveman, ultra, wenyan); scoped to specific models/providers via include/exclude filters; configurable on-error behavior (allow/block)\n- **Pricing management** — per-model pricing overrides, recalculation, import/export\n- **Usage budgets** — per-key usage tracking and limits, per-request budget enforcement via workflow feature flags\n\n### Developer Experience\n\n- **Single binary** — `npm install -g iaurora` or `docker pull aurorahq/aurora`\n- **CLI** — `aurora init`, `aurora models sync/diff/show`, `aurora update`, `aurora uninstall`\n- **CLI tools API** — admin REST endpoints for CLI configuration sync, gated separately\n- **Swagger docs** — `/swagger/index.html` (build-tag gated)\n- **Config profiles** — pre-built configs for local, local-power, and team deployments\n- **3-layer config** — code defaults → config.yaml → env vars (env vars win)\n- **Helm chart** — deploy on Kubernetes with pre-built Helm chart\n- **Docker Compose** — full infrastructure stack: Redis, PostgreSQL, Qdrant, Prometheus, Grafana\n- **Grafana dashboard** — pre-configured panels for request rate, errors, latency, in-flight requests, per-model breakdown\n\n---\n\n## Quick Start\n\nStart routing AI traffic in 60 seconds.\n\n### Option A — CLI (npm)\n\n```bash\nnpm install -g iaurora\nmkdir my-gateway && cd my-gateway\naurora init # creates config.yaml, .env, data/\n```\n\nSet your provider keys in `.env`:\n\n```env\n# ── REQUIRED ──────────────────────────────────────────────\nAURORA_MASTER_KEY=\"your-secure-key\"\n\n# ── PROVIDER API KEYS (at least one) ─────────────────────\nOPENAI_API_KEY=\"sk-...\"\nANTHROPIC_API_KEY=\"sk-ant-...\"\nGEMINI_API_KEY=\"...\"\nGROQ_API_KEY=\"gsk_...\"\nDEEPSEEK_API_KEY=\"...\"\nOPENROUTER_API_KEY=\"...\"\nXAI_API_KEY=\"...\"\nZAI_API_KEY=\"...\"\nMINIMAX_API_KEY=\"...\"\nAZURE_API_KEY=\"...\"\nORACLE_API_KEY=\"...\"\nOLLAMA_API_KEY=\"...\"\nVLLM_API_KEY=\"...\"\nJINA_API_KEY=\"...\"\n\n# ── OPTIONAL FEATURE TOGGLES (set true to enable) ────────\nLOGGING_ENABLED=true # Audit logging to storage\nMETRICS_ENABLED=true # Prometheus /metrics endpoint\nGUARDRAILS_ENABLED=true # Content safety filters\nTOKEN_SAVER_ENABLED=true # Output compression to cut token use\n\n# ── PRODUCTION STORAGE ───────────────────────────────────\n# STORAGE_TYPE=postgresql\n# POSTGRES_URL=postgres://user:pass@localhost:5432/aurora\n\n# ── REDIS CACHE (model cache + response cache) ──────────\n# REDIS_URL=redis://localhost:6379\n# RESPONSE_CACHE_SIMPLE_ENABLED=true\n```\n\n```bash\naurora\n```\n\n### Option B — inline env vars (no `.env` needed)\n\n![\"OSS](\"https://raw.githubusercontent.com/aurorallm/aurora/main/docs-assets/assets/comparison.png\")
\n\n\n![\"Email](\"https://img.shields.io/badge/Email%20Us-Enterprise-5865F2?style=for-the-badge&logo=gmail&logoColor=white\"){kind=icon}
\n \n