Your whole AI stack, under one gateway.

Two ways to scale with LunaRoute: a unified gateway over every model you use, and private inference that squeezes more throughput out of dedicated servers you control.

▸one gateway: BYO keys or use our providers
▸private inference scheduler for dedicated GPUs
▸budgets & policy by user, team, key or workload

One gateway for every model your teams already use.

Put LunaRoute in front of your provider keys or use ours. Teams get one OpenAI-compatible endpoint, finance gets one ledger, and platform engineering gets policy that follows users, keys, agents, and workloads.

▸BYO provider keys, LunaRoute-hosted providers, or both
▸budgets by user, team, key, CI job, or workload
▸failover and routing without changing agent config
▸one ledger for public and private inference spend
▸agent tools over MCP — web search, document conversion with OCR, image generation and editing — on the same key, with per-member tool policy and every call in the ledger

teams + agents  
   |  
   |  OpenAI-compatible or Anthropic-compatible requests  
   v  
+--------------------------+  
| LUNAROUTE ENTERPRISE    |  
|  |- policy + budgets     |  
|  |- routing + failover   |  
|  '- ledger + accounting  |  
+-----------+--------------+  
            |  
  +---------+---------+---------+  
  v                   v         v  
BYO keys         LunaRoute   private  
providers        providers  inference  

Private inference, without idle dedicated GPUs.

Bring dedicated capacity under the same gateway. LunaRoute can run private models on servers you control, route sensitive workloads there first, and spill only the traffic you allow.

glm-5.3            agentic coding default  
kimi-k3            long-horizon planning  
qwen-3.8-max       fast, cheap workhorse  
minimax-m3         whole-repo context  
deepseek-v4-pro    deep reasoning & review  

Single-tenant servers

Dedicated GPUs, dedicated caches, and deployment boundaries that match your risk model.

Gateway policy stays central

Keep users on one endpoint while routing sensitive workloads to your private fleet.

Utilization without routing sprawl

Keep dedicated capacity busy through the same gateway policy, ledger, and scheduler instead of sending teams to separate endpoints.

Idle GPUs are expensive GPUs.

Once the rig is a fixed cost, utilization is savings. The LunaRoute scheduler does not push past the server limit; it controls who enters the backend, keeps realtime and agent work responsive, and lets flex or background work use the gaps.

$ lunaroute-cli sched status srv-east-1  
  
srv-east-1   8 x B200   single-tenant  
admission cap ............... max_in_flight = 64  
efficient target ............ b_star = 56  
backend waiting ceiling ..... 0  
realtime / agents ........... priority 0  
flex delivery work .......... priority 10  
background work ............. priority 100  
capacity rule ............... never admit past max_in_flight  
gap fill .................... use spare seats, yield on deadline risk  

▸ Priority classes
Realtime and agent calls get priority; flex and background work wait behind them when deadlines compete.

▸ Headroom admission
Normal work enters only while the backend has headroom; no admission crosses the hard max_in_flight cap.

▸ Queue with backpressure
Excess work waits in the scheduler queue or fails fast when the configured deadline can no longer be met.

▸ Gap-filling work
Flex and background jobs can fill spare slots between realtime bursts, raising useful work without pretending the server has more capacity.

NO LOGS, at enterprise scale.

The same posture as the public product, contractually: request and response bodies are processed in memory and never stored by LunaRoute. On private inference, traffic stays on your single-tenant path.

No request-body storage

Prompts and completions are never written to disk by LunaRoute. We keep timestamps, model, token counts and cost - your ledger.

Single-tenant isolation

Dedicated servers are yours alone - no shared GPUs, no shared caches, no co-mingled traffic.