Skip to main content

The LLM gateway that keeps native power.

One Rust binary. OpenAI-compatible on the outside, native Vertex AI and Jev routing on the inside. No lowest-common-denominator adapters.

Get startedStar on GitHub
docker run --rm -p 8080:8080 -p 9090:9090 \
  -e VERTEX_PROJECT_ID=my-gcp-project \
  -e OPENAI_API_KEY=sk-... \
  -e GOOGLE_APPLICATION_CREDENTIALS=/secrets/sa.json \
  -v "$(pwd)/sa.json:/secrets/sa.json:ro" \
  -v "$(pwd)/config:/app/config" \
  sustentabilitas/synapse-gateway

# In a second terminal:
curl -s http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-flash",
    "messages": [{"role": "user", "content": "Describe this video."}],
    "vertex": {
      "media_uris": ["gs://cloud-samples-data/video/animals.mp4"]
    }
  }'

Native first

Keep the capabilities other gateways throw away

Most gateways reach every provider through one generic adapter. Synapse keeps dedicated native lanes, so provider-specific features survive the trip.

Native Vertex AI lane

Context caching, gs:// media URIs, strict responseSchema decoding and thinking config go straight to Vertex AI streamGenerateContent. Nothing is flattened through an OpenAI-shaped adapter.

Native tool calling

Tools work on the standard and native Vertex lanes, and on the native lane tool_choice maps to a Vertex toolConfig mode instead of being dropped.

Jev lane

Ask TypeSafe System One typed questions and get structured decisions back, or judge candidates and extract only the survivors in one call.

Jev router

Jev rates each request’s difficulty and serves it from the matching model tier and reasoning effort. Effort becomes each provider’s own control, such as a Vertex thinking budget, and every response reports the decision in x-synapse-* headers.

Real streaming, full fallback

On the standard and native Vertex lanes, Synapse always streams from upstream. Streaming clients get token-by-token SSE; buffered standard-lane clients keep the whole fallback chain, even after a mid-stream failure.

Embeddable

Run one small Rust binary, or drop the library crate into your service and call Gateway::chat() in-process.

And everything you expect

Architecture

Three lanes, one OpenAI-compatible front door

Every chat request passes guardrails, then lane detection picks the standard, native Vertex or Jev lane. Each route is an ordered fallback chain, and every call lands in the cost ledger and the metrics pipeline. Read the architecture guide

ClientGuardrailsLane detectionStandard laneOpenAI · QwenVertex · vLLMNative Vertex laneVertex AIJev laneTypeSafeFallback chains · Cost ledger · OpenTelemetry metrics

The Synapse family

More than a gateway

synapse-proxy

A config-driven reverse-proxy sidecar: path-prefix routing, context injection, request and response transforms, streaming passthrough.

synapse-a2a

An in-memory A2A agent registry on the gateway’s API port: register agents, then let clients discover them, fetch their agent cards and resolve their URLs.

synapse-mcp

An MCP gateway that routes tool calls to servers registered on demand and injects the caller’s tenant identity, so tools see who is calling.

Open source

Open, free and built in the open

Synapse is licensed under MPL-2.0, so you can build it into commercial products. Every feature on this page is in the open-source code, with no paid tier, and improvements contributed back are welcome.

GitHubContributecrates.ioDocker HubBlog