Give every request
the gateway it deserves
Synapass is the OpenAI-compatible AI gateway that routes, caches, and governs multi-provider inference —
and proves every decision it makes.
Overview
Traffic, spend and health for the selected window.
Requests
1.28M
+8.2%
Success rate
99.98%
+0.1%
Latency p95
1.9s
−10.5%
Spend
$41.27
+12.4%
API requests
Last 24 hours · 1,284,092 total · UTC
Synapass control plane · illustrative preview — your live data appears after install
Speaks OpenAI · Routes to everyone
Nine steps. Every one recorded.
Authenticate, resolve policy, enforce budgets, check cache, classify the task, shape the prompt, route, execute with failover, and run the tools loop. Select any step to read its explain detail.
key syn_live_… → tenant acme · scope inference
P95
1.9s
Success
99.98%
Spend today
41.27$
Restarts
00to re-route
Illustrative figures — yours appear after install.
Five patterns. One always fits.
Routing strategies are panel patterns: pick one per policy, switch without deploying, watch the bank re-balance in analytics.
Routing strategies
05 patterns
- 01PriorityDefaultAn ordered list; the first healthy provider answers. Best for a clear first choice with ordered fallback.first healthywins
- 02WeightedTraffic split by proportion across eligible providers. Best for canarying a new model and A/B comparisons.by weightsplit
- 03Lowest costThe cheapest eligible model serves each request. Best for batch, background, and cost-capped tenants.min priceper request
- 04Lowest latencyRoutes on observed response times, not marketing numbers. Best for interactive paths where milliseconds show.min msobserved
- 05Highest qualityRoutes on observed scores from evals and replay. Best for correctness-first work like review and summary.max scoreobserved
Every provider, one bus.
OpenAI, Anthropic, Ollama, vLLM, and any OpenAI-compatible API patch into the same bus. Kill or revive a channel live — traffic reroutes, nothing restarts.
Providers
Live bus
- openai-prodgpt-4o-mini812 ms62% bus
- anthropic-euclaude-3-5-haiku1.1 s27% bus
- local-ollamallama3.1:8b340 ms11% bus
- legacy-vllmmistral-7b—0% bus
Illustrative bus levels — your providers patch in after install.
Running in about five minutes.
One command clones, configures, and starts the full stack — gateway, console, PostgreSQL, Redis, ClickHouse, NATS, workers, and observability.
Start the stack
curl -fsSL https://raw.githubusercontent.com/shadowsafin/synapass/main/scripts/install.sh | bashOne command clones, configures, and starts everything — re-runs are safe.
Docker guideMint a tenant key
TENANT=$(curl -s $GATEWAY/admin/v1/tenants -H "Authorization: Bearer $SYNAPASS_ADMIN_KEY" | jq -r '.tenants[0].id')Plaintext is returned exactly once and stored only as a digest.
Getting startedMake the call
curl -s $GATEWAY/v1/chat/completions -H "Authorization: Bearer $SYNAPASS_KEY" -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"hello"}]}'Any OpenAI client works — change the base URL and go.
API reference
curl -fsSL https://raw.githubusercontent.com/shadowsafin/synapass/main/scripts/install.sh | bashSafe by construction.
A gateway that holds provider credentials and customer prompts must fail closed. Production validation refuses unsafe config before serving a single request.
- Tenant isolation, structuralRouting, cache keys, and namespaces isolate tenants by construction — a hit never crosses tenants.3 layersisolated
- Secrets referenced, never storedCredentials name environment variables or are sealed with AES-256-GCM. Inline secrets are discarded.0 inlinesecrets kept
- Fails closed in productionCORS wildcards, missing admin keys, and dangling tracing refuse to boot — loudly, before serving.3 checksenforced
- Every mutation auditedControl-plane changes record before and after values. Inference stays fast; usage records serve history.100%of mutations
Put Synapass on the bench tonight.
One command starts the full stack. Your first routed call lands in about five minutes — with the lineage to prove it.