Open source · AGPL-3.0Go 1.25+

One gateway.
Every LLM dialect.

Inja sits between your agents and providers. Clients keep talking OpenAI, Anthropic, or Gemini — you route, translate, and meter without rewriting every integration.

Get started How it works GitHub

Single static binary
yaml.v3 only at runtime
No vendor SDK lock-in
AGPL-3.0

gateway.yaml

# multi-provider edge
providers:
  openai:
    kind: openai
    base_url: https://api.openai.com/v1
  anthropic:
    kind: anthropic
    base_url: https://api.anthropic.com/v1
  deepseek:
    kind: openai_compat
    base_url: https://api.deepseek.com

defaults:
  openai_dialect: openai
  anthropic_dialect: anthropic

aliases:
  fast: deepseek/deepseek-chat

hooks:
  jsonl:
    output: stdout

Request path

Clients stay native. Upstreams stay flexible.

A thin edge process — YAML config, no database, no vendor SDKs at runtime. Same-family traffic never pays a translation tax.

Ingress

POST /v1/chat/completions
POST /v1/messages
…/models/*:generateContent
/v1/responses · files · batches

Gateway

01Resolve alias / provider / model
02Auth edge keys + upstream api_key_env / ADC
03PT or IR passthrough vs translate
04Hooks usage JSONL / webhook

Egress kinds

openaifirst-party
openai_compatany host
anthropicMessages
googleGemini native

How it works

Four moves on every request

01
Keep your client dialect
OpenAI Chat Completions, Anthropic Messages, or native Gemini — same edge, same auth story.
02
Route by model id
Bare models hit defaults. provider/model prefixes and aliases pick the upstream — DeepSeek, xAI, Groq, OpenRouter, vLLM, …
03
Passthrough or translate
Same-family traffic is byte-faithful. Cross-family rebuilds through a canonical IR and fails closed on unsupported shapes.
04
Meter every request
One usage event per call (JSONL / webhook): model, provider, tokens, cache, latency, status.

Deep dive: request path

Built for agents

Production surfaces, not demos

Fidelity policies, platform API proxies, and operator hooks — documented and hermetically tested.

Claude Code, Claude app, Codex, Cursor

ChatGPT, Claude, and SuperGrok via OAuth — same gateway from Claude Code, the Claude desktop app, Codex, or Cursor.

Prompt cache policy

Preserve cache_control and prompt_cache_* within family. Opt-in auto breakpoints for OpenAI/Google → Anthropic.

Platform proxies

Files, Batches, Assistants, Skills, tuned models, File Search, cachedContents — thin upstream-owned surfaces.

Operator hooks

Prometheus /metrics, optional provider health, edge auth, body limits, and observe_dropped_fields.

Quick start

Ship in minutes

Grab a release binary or run Docker. Drop a YAML file. Point your SDK at the edge.

Install guide Compatibility matrix

dockerbinary

# compose up with your keys
docker compose up -d

# or grab the release binary
curl -sL https://github.com/inja-online/llm-gateway/releases/latest

# health check
curl -s localhost:8080/healthz

Multi-provider without multi-SDK pain

Install once. Put OpenAI-, Anthropic-, or Gemini-shaped clients behind one edge — and swap backends with config, not code.

Read getting started Star on GitHub

One gateway.Every LLM dialect.