RelayGate is programmable middleware for AI traffic. One statically-compiled Go binary, CGO=0, no runtime dependencies. Three inbound wire formats translate to one canonical ChatRequest. Ten backend drivers. CEL for everything.
Other AI gateways put your middleware in front of or behind the request. RelayGate runs it inside. ContextWorkers mutate, enrich, block, or audit against the parsed envelope at sub-millisecond overhead. That inline capability is the wedge.
| id | Name | Kind | Format | Notes |
|---|---|---|---|---|
openai | OpenAI | direct | openai | GPT series, o-series |
anthropic-direct | Anthropic | direct | anthropic-messages | Claude Opus, Sonnet, Haiku |
google-gemini | Google Gemini | direct | gemini | Gemini Pro, Flash |
groq | Groq | direct | openai | LPU-accelerated Llama, Mixtral |
deepseek | DeepSeek | direct | openai | V-series, R-series |
together | Together AI | direct | openai | Open-weight hosted models |
mistral | Mistral | direct | openai | Ministral, Mistral Large |
cohere | Cohere | direct | cohere | Command R, Command R+ |
openrouter | OpenRouter | aggregator | openai | Aggregated access to 200+ models |
local-ollama | Local (Ollama) | direct | openai | Local models over OpenAI-compatible endpoint |
The core is Apache 2.0. The drivers are MIT. The managed tier is a commercial service layered on top of the same open binary. You can run every feature of RelayGate without paying us a cent. Self-host, full feature set, community support. That is the deal.
RelayGate composes with adjacent infrastructure when you want it to:
Every integration in RelayGate is optional. Run RelayGate by itself and it still delivers the wedge. Pair with R1 for inline agents, DeepTap for grounding, TrueCom for settlement, RelayOne for fleet governance. None is required.
Sales, partnerships, or anything else. Mail lands at [email protected].