Your infrastructure. Your keys. Your contracts.
RealityRouter is MIT-licensed and runs wherever you run — your data centre, your VPC, or your own cloud account. Prompts go straight from your infrastructure to your providers using your own keys, and at Enterprise the calibration service can run inside your perimeter too.
Self-hosted or hosted — your choice
The router already runs on your own infrastructure at every tier. At Enterprise, RealitySignal calibration can too.
Nothing leaves your perimeter
RealitySignal runs in your VPC or on-prem. Prompts, completions, the decision log, and calibration all stay inside — we never see any of it. Built for teams that can’t let data leave: banks, hospitals, research institutions.
Same cloud, Enterprise support
RealitySignal runs in our cloud, same as Team — still only anonymised task features, never prompt text. Built for teams whose security posture already approves standard cloud deployments and just need identity, audit, and an SLA on top.
The same, either way
Your keys, your provider contracts.
If you’ve negotiated rates or committed spend with a model provider, routing through a reseller means paying their prices, not the ones you agreed. RealityRouter uses your own keys and calls your providers directly, so your discounts keep applying and your relationship stays yours.
No percentage of your inference spend.
Routers that sit in the request path typically take a cut — 5.5% is the going rate. At $100,000 a month of inference that is $5,500 a month, and it grows precisely as you succeed. We charge for calibration. Your model bill is not our revenue.
MIT, so there is no rug to pull.
The router is open source and runs on your infrastructure. If we raise prices, change direction, get acquired or simply disappear, you keep routing. A closed hosted router is a dependency you cannot replace in an afternoon, and procurement increasingly asks what happens if the vendor goes away.
What’s included
Everything in Team, plus:
- Self-hosted (VPC or on-prem) or hosted — your choice
- SSO / SAML
- Role-based access and an exportable audit log
- DPA and data residency commitments
- SLA with a named contact and a shared channel
- Help getting set up: model pool design and calibration tuning on your own traffic
Why calibration is the part that is hard to copy
Several good decision models shipped this year — some with open weights you can run locally in milliseconds. Use them. RealityRouter will route to them.
What none of them do is calibrate to your traffic. They are trained once and shipped, and their published accuracy is measured on a public benchmark, not on your prompts, your model pool, or your definition of a good outcome.
RealitySignal is not a model. It is a statistical calibration layer with a mathematical guarantee attached, trained on your own observed results and tightening as evidence accumulates. It can sit on top of whichever decision model you prefer, including the open ones.
That is the difference between a model calibrated once, somewhere else, and calibration that is true where you are running.
Questions people ask
- Do you see our prompts?
- No. On every tier, prompts go from your infrastructure straight to your provider. Our hosted calibration service receives anonymised task features only. Choose self-hosted Enterprise and even that stays inside your network.
- What happens if we stop paying?
- The router is MIT. It keeps working, forever, with no calibration service. You are not locked in, by construction.
- Can we audit it?
- Yes. The router source is public, the routing maths is documented, and the decision log is yours.
- Can we use our own models?
- Yes — any OpenAI-compatible endpoint, including local models and anything you host yourself.
- Do you support air-gapped deployment?
- Talk to us. Self-hosted RealitySignal is in production; a fully air-gapped install depends on your constraints.