Every model your company uses, on one balance you can cap
No spread of ours on the model price, one percentage at top-up, a spending ceiling the gateway enforces before it calls anyone, and a ledger that reconstructs every charge. That is the offer. What we have not built is written down further down this page, in full, so that nobody plans around it.
- 0%
- spread of ours on model prices
- billed at what the route costs us, per token
- 5%
- our entire margin, taken at top-up
- plus $0.35 card processing at cost; both 0% on your own keys
- 402
- returned instead of an overdraft
- ceilings are checked before any upstream call
- 100%
- of charges reconstructable from the ledger
- append-only; npm run db:audit proves it
Talk to an engineer
There is no qualification form and no sales team to route one to. Messages go to the person who wrote the gateway — and if the answer is that Multigrid does not fit what you need, that is the answer you will get.
Send a message. No account needed.
There is no address printed here because our mailbox is not receiving mail yet — it would bounce at the sending server and you would never know. The contact form writes to a database we own, which is the only route we can prove reaches us. A page that lists an address which silently bounces is exactly the sort of thing this site has been getting rid of.
Four lines that make the first reply useful:
- Roughly what you spend on models a month, and with whom.
- The shape of the traffic — chat, batch, agents, embeddings.
- Any constraint that is non-negotiable for you: residency, an audit report, SSO. Lead with it; it may end the conversation early, which is better for both of us.
- Whether you hold provider contracts you would rather keep using.
No response time is promised here, for the same reason there is no SLA below: there is no rota behind it yet.
Six things you can verify in an afternoon
Every one of these is running today. Fund an account with twenty dollars and you can check the lot against your own traffic before you talk to anybody here.
One balance, no spread of ours, one number to explain
Every model is charged at what its route costs us — a direct provider's published rate, or an OpenRouter rate carrying their 5.5% credit fee, which the model page names either way. Our margin is a single percentage taken when you add credit, itemised on the receipt, and there is nothing else. Finance can reconcile a month from the ledger without asking us what a line means.
A ceiling that actually stops
Set a monthly cap on the account and a separate cap on each API key. Both are enforced in the gateway before it calls any provider: past the limit a request is refused with a 402 and a code naming the ceiling it hit. No overdraft, and no grace spending to argue about afterwards.
Your provider contracts, our routing
Add your own OpenAI, Anthropic or other keys and we route on your rates. Keys are encrypted with AES-256-GCM, never returned to any client, and only their last four digits are ever displayed. Traffic served on your keys costs you nothing from us at all.
Failover across the providers serving a model
Where more than one provider serves the same model, a failing one is stepped over automatically and the retry is not billed twice. Pin a single provider per request when you would rather fail than be moved — a pin is treated as an instruction, not a preference.
Guardrails, and a log that holds no prompts
Refuse requests carrying PII, terms you supply, or more tokens than you allow — checked before the request is priced. The request log records tokens, latency, cost, model, provider and our routing trace, and no message content.
Into your observability, not another dashboard
Export one OTLP span per request to Tempo, Honeycomb, Datadog or your own collector, so the latency we add sits inside the trace of the request that was waiting on us. Signed webhooks cover the money events: balance low, balance depleted, key cap reached, spend threshold, request failed.
The list your security review is looking for
This page used to advertise every item below as shipped. None of it is. It is printed here in full so that nobody discovers it in week three of a pilot.
No SOC 2, ISO 27001 or completed CAIQ
No audit has been performed and no report exists. If you have been told otherwise, you were told wrong. We will answer a security questionnaire honestly, which for several rows means writing “no”.
No SSO or SCIM
Sign-in is email and password. There is no SAML, no OIDC and no directory provisioning. Password reset exists but cannot deliver its mail until a provider is connected, so account recovery currently goes through support by hand. Projects, roles and per-project access do work.
No data residency guarantees
There is no region pinning and nothing that fails closed at a border. A request goes to the provider endpoint that serves the model.
No private or self-hosted deployment
There is one managed multi-tenant service. No single-tenant VPC, no on-prem binary, no air-gapped mode.
No SLA and no support rota
We will not sell you 99.99% with service credits or a one-hour response target, because nobody is on a pager to honour it. Live provider status is published and incidents are not smoothed over. There is no support mailbox either — the domain has no mail routing yet, so the only channel that provably arrives is the support form inside the product.
No usage history to point at
This has only just launched. There is no trailing-twelve-months availability figure, no customer list and no case study — and saying so is better than inventing one.
If one of these is a hard requirement, Multigrid is not the right choice for you this year and no amount of conversation changes that. Tell us which one anyway: a requirement we hear from a real buyer is worth more to us than a pilot that was never going to close.
What procurement asks
What does it cost at volume?
The 5% top-up percentage is negotiable downwards at volume, and that is the entire commercial conversation: there is no licence, no seat count and no feature locked behind a tier. A flat $0.35 of card processing sits beside it on each top-up, passed through at cost and irrelevant at any volume worth discussing. Traffic on your own provider keys is 0% regardless of volume, so for a team that already holds provider contracts the honest answer is often that we should be charging you very little.
How do we evaluate this without betting anything on it?
Fund an account with a small amount of credit and set the monthly ceiling to that same amount. The cap is enforced before every upstream call, so the worst possible outcome of a bad week is the number you funded. Point one non-critical workload at the base URL — it speaks the OpenAI wire format, so that is a config change rather than a migration — and compare the per-request cost in Activity against what you run today.
What happens to our data?
The request log stores token counts, latency, cost, the model and provider, and our routing trace. It does not store your messages. Two things do hold content, because they cannot work otherwise: a cached response, keyed per account so it can never be served to another customer, and a batch job, which keeps each submitted line and its answer until the job is deleted. Trace export carries counts and timings only. We do not train on any of it and we do not sell any of it.
What if we outgrow you, or you disappear?
You are calling an OpenAI-compatible endpoint, so leaving is the same config change as arriving, pointed somewhere else. Credit is refundable minus payment processing. And if you are running on your own provider keys, your contracts are with those providers and survive anything that happens to us — which is most of why bring-your-own-key exists.
Can we get an invoice instead of paying by card?
Ask. Card top-up through Stripe is what is built and automated today; anything else is a manual arrangement, so it depends on the size, and we would rather agree it with you than advertise a billing process we have never run.
The full price list is on pricing, and live provider availability is on status.