Skip to content

Repository files navigation

AISIX AI Gateway

The open-source, Rust-native AI gateway for LLMs and AI agents

One OpenAI-compatible API in front of every model. Route, govern, secure, cache, and observe all your LLM and AI-agent traffic from a single control point — shipped as one static binary with low per-request overhead. Self-host for free, forever.

Built by the original creators of Apache APISIX.

License: Apache 2.0 Built with Rust Docs Discord Website

Start free · Documentation · Quickstart · AISIX Cloud · Roadmap


AISIX AI Gateway architecture — one OpenAI- or Anthropic-compatible API in front of OpenAI, Anthropic, Gemini/Vertex, Bedrock, Azure OpenAI, and DeepSeek, with API key auth, rate and token limits, guardrails, caching, routing and failover, and observability in between

AISIX AI Gateway is a Rust-native gateway that puts a single, OpenAI-compatible API in front of every LLM provider — OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, DeepSeek, and any OpenAI-compatible endpoint. It gives platform teams one place to route, govern, secure, and observe LLM traffic, with first-class SSE streaming and low gateway overhead.

It runs as a single static binary — low cold-start, lock-free config reads, and hot configuration reloads with no restarts: declare resources in one resources.yaml and reload on SIGHUP, or point the gateway at etcd for a multi-replica cluster. Run it self-hosted and free, or connect it to AISIX Cloud for a managed control plane with team governance, budgets, audit, and a dashboard.

AISIX AI Gateway (this repo) is the open-source core — the gateway, i.e. the data plane. AISIX Cloud adds a managed control plane on top, either operated by API7 (Hybrid Cloud) or run entirely in your own infrastructure (On-Premises). In every form the gateway runs in your environment and calls providers directly — live AI traffic never leaves your infrastructure. The proxy API is identical throughout. Talk to us about AISIX Cloud →

⚡ Quickstart

One container. No control plane, no database, no configuration store — the gateway reads every dynamic resource from one declarative resources.yaml.

# config.yaml
resources_file: /etc/aisix/resources.yaml
proxy:
  addr: "0.0.0.0:3000"
admin:
  enabled: false          # a declarative gateway needs no admin listener
observability:
  metrics:
    prometheus:
      enabled: true
      addr: "0.0.0.0:9090"
# resources.yaml
_format_version: "1"

provider_keys:
  - display_name: openai-main
    provider: openai
    api_key: ${OPENAI_API_KEY}        # interpolated from the environment

models:
  - display_name: my-model
    provider: openai
    model_name: gpt-4o-mini
    provider_key: openai-main

api_keys:
  - display_name: local-dev
    key_env: CALLER_API_KEY           # hashed at load; the plaintext is never stored
    allowed_models: ["my-model"]
export OPENAI_API_KEY="YOUR_PROVIDER_KEY"
export CALLER_API_KEY="YOUR_CALLER_KEY"

docker run -d --name aisix \
  --platform linux/amd64 \
  -v "$(pwd)/config.yaml:/etc/aisix/config.yaml:ro" \
  -v "$(pwd)/resources.yaml:/etc/aisix/resources.yaml:ro" \
  -e OPENAI_API_KEY -e CALLER_API_KEY \
  -p 3000:3000 -p 127.0.0.1:9090:9090 \
  ghcr.io/api7/aisix:latest        # proxy → :3000, metrics + status → :9090
#                                  ^ the metrics/status listener is unauthenticated;
#                                    keep it on loopback or a private network

Then call the gateway exactly like OpenAI:

curl http://localhost:3000/v1/chat/completions \
  -H "Authorization: Bearer $CALLER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"my-model","messages":[{"role":"user","content":"hello"}]}'

Edit resources.yaml and send SIGHUP (docker kill -s HUP aisix) to apply changes with no restart — an invalid file is rejected whole and the last good configuration keeps serving. Check a file before booting with aisix validate --resources resources.yaml.

Full walkthrough: the Gateway Quickstart · every field: the resources file reference. For a multi-replica cluster, point the gateway at etcd instead — resources_file and etcd are mutually exclusive.

✨ Why AISIX

  • One API, every model. Speak the OpenAI or Anthropic wire format in; the gateway translates to whichever provider each model points at. Point an OpenAI or Claude SDK at one base_url and switch models without changing code.
  • A real gateway, in Rust. Single static binary, low cold-start, lock-free config reads on the hot path, native streaming.
  • Open-source core, free forever. Apache-2.0, self-hostable end to end. Reach for AISIX Cloud only when you want the managed control plane.
  • Production controls built in. Routing & failover, rate limits, guardrails, caching, and observability ship in the box. (Budgets and spend caps are an AISIX Cloud feature — the gateway enforces the control plane's decisions.)

🧩 Features — available today

Covered by 165 end-to-end test files (426 cases) that run against real gateway processes.

  • OpenAI-compatible proxy (:3000) — chat/completions, completions, responses, embeddings, rerank, images/generations, audio/{speech,transcriptions,translations}, videos (submit → poll → fetch), files, batches, fine_tuning/jobs, realtime, GET /v1/models, plus a root-level /passthrough/:provider/* escape hatch. Native SSE streaming, tool/function calling, JSON mode, vision/multimodal input, and reasoning-content support.
  • Anthropic Messages APIPOST /v1/messages as a first-class route, working against any configured upstream: requests and responses (including streaming) are translated both ways when a model points at a non-Anthropic provider.
  • Routing & failover — virtual/routing models with six strategies: round_robin, weighted (with sticky/canary hashing), failover, plus metric-based least_cost, least_latency, and least_busy. Retry budgets, cooldowns, tag-conditional targets, and per-attempt timeouts.
  • Ensemble models — fan one request out to a panel of models concurrently, then have a judge model synthesize a single answer, with a minimum-successful-responses threshold.
  • Semantic routing — one virtual model that dispatches by the meaning of each request: it embeds the prompt, scores it against per-route example utterances, and routes to the best match (or a default). See the semantic routing docs.
  • Rate limiting & concurrency — RPS/RPM/RPH/RPD + TPM/TPD + concurrency caps, AND-combined across caller keys, models, and policy scopes (api_key / model / team / member / team_member). Counters are per-process by default, or shared across replicas with the Redis backend.
  • Guardrails — content-policy enforcement on input and output, in-process or through a provider: keyword/regex, built-in PII detection and redaction, Presidio, Lakera, OpenAI Moderation, AWS Bedrock Guardrails, Azure AI Content Safety (Prompt Shield + text moderation), and two Alibaba Cloud services. A block returns 422 content_filter; monitor mode records what would have happened without blocking.
  • Caching — exact-match response cache with per-policy TTL and model/key scope matchers; memory and Redis backends; cost-saved telemetry on every hit. Separately, automatic prompt caching can be enabled per direct Anthropic model to inject cache breakpoints, so callers get provider-side prompt discounts without changing their requests.
  • MCP gateway — register upstream MCP servers as first-class resources and front them all behind one endpoint (/mcp), with the gateway holding each server's upstream credential, namespacing tools per server, and enforcing per-caller tool access. Also exposes a REST API as MCP tools from its OpenAPI description.
  • A2A agent gateway — front A2A (Agent-to-Agent) agents at /a2a/:agent, serving each agent's card with URLs rewritten to the gateway, over JSON-RPC 2.0.
  • Inbound authentication — caller API keys (SHA-256 hashed, model allowlists, expiry, rotation), or OIDC/JWT bearer tokens validated against registered providers (Entra ID, Okta, Google Workspace, or any OIDC issuer) with JWKS caching.
  • Observability — Prometheus /metrics, structured per-request access logs, usage events, OTLP/GenAI span export (Langfuse, Honeycomb, Grafana Cloud, or any OTLP receiver), plus dedicated Datadog and Aliyun SLS log exporters and object-storage (S3/GCS/Azure Blob) telemetry.
  • Declarative configuration — one resources.yaml carries all ten resource collections (provider keys, models, caller keys, guardrails, MCP servers, A2A agents, cache policies, observability exporters, rate-limit policies, OIDC providers), validated against the same JSON Schemas the gateway uses at runtime. aisix validate checks a file offline; SIGHUP reloads it atomically.
  • Operational endpoints/livez and /readyz on the proxy listener; /status/config, /status/ready, /status/models, and Prometheus /metrics on a dedicated metrics listener (:9090). The admin listener (:3001) additionally serves a resource read surface, OpenAPI 3 with a Scalar UI, and a playground. Its resource write endpoints are deprecated in favor of declarative configuration: on a resources_file gateway they are rejected with 409, and on a store-backed gateway they still work but every mutating response carries an RFC 9745 Deprecation header.

🔌 Supported providers

AISIX dispatches through five native adapter families — distinct wire-protocol bridges, not one generic relabel. Whatever the upstream protocol, the client-facing API stays OpenAI-shaped.

Adapter family Reaches Wire shape · auth
openai OpenAI + any OpenAI-compatible vendor — DeepSeek, Groq, Mistral, Together, Fireworks, Perplexity, vLLM, Ollama, self-hosted OpenAI chat completions · Bearer
anthropic Anthropic Claude Anthropic Messages · x-api-key
bedrock AWS Bedrock — Anthropic, Meta Llama, Mistral, Cohere, Amazon Titan/Nova, AI21 Bedrock Converse + /invoke · SigV4
vertex Google Vertex AI (Gemini) Vertex :generateContent · OAuth2
azure-openai Azure OpenAI Azure deployments · api-key / Entra ID

Plus specialized handling for vendor quirks (e.g. DeepSeek reasoning content) and dedicated rerank / embeddings vendors (Cohere, Jina). Details in adapter protocol families.

☁️ Open source vs AISIX Cloud

Same gateway binary, same proxy API — in every form the gateway runs in your environment. AISIX Cloud adds a managed control plane on top, either operated by API7 (Hybrid Cloud) or running entirely in your infrastructure (On-Premises).

AISIX Cloud overview — requests, latency p50/p99, error rate and cost today, with a 7-day request-and-cost trend and data-plane health
Overview — traffic, latency, error rate & spend at a glance

AISIX Cloud models — alias an upstream LLM per provider (OpenAI, Anthropic, AWS Bedrock, DeepSeek) with model IDs and per-model rate limits
Models — one alias per upstream: OpenAI, Anthropic, Bedrock, DeepSeek…

AISIX Cloud guardrails — pre-input and post-output content policies (keyword blocklist, Azure Content Safety, AWS Bedrock) that block on violation
Guardrails — pre-input & post-output policies, block on violation
AISIX Cloud playground — pick a model, set system and user prompts, run, and read the response with live token and cost metering
Playground — test any model with live token & cost metering

AISIX Cloud observability exporters — fan out chat-completion telemetry to OTLP, Datadog and object storage, with per-target delivery health
Observability — fan out traces & logs to OTLP, Datadog, object storage

AISIX Cloud budgets — organization and per-environment spend caps with progress bars, hard-stop versus warn-only, including an over-budget policy
Budgets — hard-stop spend caps with warn-only tiers

The AISIX Cloud dashboard — overview metrics, multi-provider models, guardrails, budgets (with hard-stop spend caps), and observability exporters, across all your gateways.

▶ Try the live dashboard demo — aisix-demo.api7.ai

Open-source gateway (this repo) AISIX Cloud (Hybrid Cloud or On-Premises)
Price Free · Apache-2.0 · forever Commercial — talk to us
Configuration Declarative resources.yaml, or etcd for a cluster Dashboard + Cloud Admin API, multi-environment
Tenancy Single instance / namespace Org → Team → Member → Environment
Provider keys In the resources file as ${VAR} env references, or in etcd Envelope-encrypted at rest, write-only, in-place rotation
Inbound auth Caller keys (SHA-256 hashed, model allowlists, expiry), or OIDC/JWT bearers Same, plus masked reveal, key ownership, and PATs
Budgets — (rate and token limits only) Per key / provider / env / org / team, hard-stop & alerts
RBAC Admin key = full access Org roles (owner / admin / member), invites
Audit log Full org-scoped audit with diff viewer
Usage & cost Export logs, metrics, and usage events yourself Managed usage views, model pricing catalog, spend reporting
Surface Status endpoints, OpenAPI read surface, playground Full dashboard + per-environment playground

Want the managed control plane, governance, budgets, and dashboard? Talk to API7 about Hybrid Cloud or On-Premises, or book a demo.

🏗️ Architecture

A single Cargo workspace; the aisix-server crate builds one binary named aisix that wires the crates together.

crates/
├── aisix-core           Config, snapshot, resource model, resources.yaml source, errors
├── aisix-etcd           Config provider + watch supervisor
├── aisix-gateway        Hub & bridge, SSE parser, provider trait
├── aisix-proxy          /v1/*, /mcp, /a2a handlers, routing, middleware
├── aisix-admin          Read surface + playground + OpenAPI (writes deprecated)
├── aisix-provider-*     openai · anthropic · azure-openai · bedrock · vertex
├── aisix-mcp            MCP gateway — server registry, tool ACL, transports
├── aisix-a2a            A2A agent gateway — agent cards, JSON-RPC bridge
├── aisix-ratelimit      fixed-window + token accounting + concurrency (local | redis)
├── aisix-cache          memory + redis backends
├── aisix-redis          shared Redis connection for cache + rate limits
├── aisix-guardrails     pre/post content-policy hooks
├── aisix-obs            tracing, metrics, access log, exporters
└── aisix-server         the `aisix` binary — bootstrap + CLI

🗺️ Roadmap

Highlights on the roadmap; tracked live in issues:

  • Semantic (embedding-similarity) response caching
  • More observability sinks — Langsmith, Helicone, Slack alerts
  • Prompt templates managed as gateway resources
  • Llama-Guard as a guardrail provider

Shipped since this list was last written: the MCP gateway, the A2A agent gateway, OIDC/JWT inbound auth, Redis-backed distributed rate limiting, and the Lakera, Presidio, PII, and OpenAI Moderation guardrails — see Features above.

🛠️ Development

Prerequisites: the Rust toolchain pinned in rust-toolchain.toml. Docker is only needed for the tests that exercise etcd, Redis, or provider emulators.

cargo check --workspace
cargo fmt --check
cargo clippy --workspace -- -D warnings
cargo test --workspace

# Coverage (matches the CI gate)
cargo llvm-cov --workspace --lcov --output-path lcov.info

# Run locally against a resources.yaml (no etcd needed). Copy the Quickstart's two files
# and change resources_file to the local path, e.g. resources_file: ./resources.yaml
cargo run -p aisix-server --bin aisix -- --config config.local.yaml

# Check a resources file without starting a listener
cargo run -p aisix-server --bin aisix -- validate --resources resources.yaml

💬 Community

If AISIX is useful to you, a ⭐ helps other engineers find it.

📄 License

Apache 2.0.

About

Open-source AI gateway for LLMs & AI agents, built in Rust. One OpenAI-compatible API for OpenAI, Anthropic, Gemini, Bedrock & more — routing, guardrails, caching, rate limits, observability.

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages