Spend control
One invoice, no idea which team spent it
Most AI budgets are discovered, not managed: the invoice arrives, someone asks who spent it, and the answer takes a week of log spelunking. We run the opposite pattern on our own spend — a cost written per call and a daily cap that refuses the next one — and we will show you exactly how far that goes.
Signs this is you
- Finance asks which product line drove last month's AI spend, and engineering cannot answer confidently.
- A retry loop or a runaway agent can burn a meaningful share of the month before anyone notices.
- Every workflow calls the strongest model available, including the ones classifying short strings.
- Nobody knows what a feature costs per user, so nobody can price it.
What we run, and what we do not
Every capability on this page carries its real status: shipped code, work we do inside a pilot, or a design we have not built yet.
A daily cap that actually refuses calls
Live todayOur router sums the day's logged spend before each call and raises rather than proceeding once the configured USD cap is reached. There is a test for exactly that behaviour, because a budget guard nobody tested is a decoration.
Cost recorded per call, not per invoice
Live todayEach call writes provider, model, operation, tokens, latency and an estimated cost in USD. Attribution starts as a real row, not a month-end reconciliation.
Provider choice per operation
Live todayThe router resolves provider and model per operation across several vendors, so cheap work does not have to run on the expensive model. The choice is recorded with the call.
Team and project attribution, caps per lane
Design intentSplitting spend by team and project, and capping a specific lane, is the design. The fields and the multi-tenant gateway are not built — today the cap is global to our own router.
How a conversation goes
1. Tell us your current setup
Live todayProviders, rough monthly spend, which workflows you suspect. Most teams already have a hunch, and it is usually wrong in an interesting way.
2. We show what the record shape gives you
Live todayWhat a per-call cost row answers, and what it does not answer without team and project fields.
3. Decide whether it is worth changing anything
Pilot scopeIf your spend is small or your stack already reports per-feature cost, we will say so. Selling a control plane to a team with one workflow helps nobody.
What exists today
- A configurable daily USD cap enforced before each model call, with test coverage — running on our own router.
- Per-call records with provider, model, operation, tokens, latency and estimated cost.
- Routing across several providers, with the selected model recorded on the call.
What we do not claim
- We publish no savings percentage. Any number would depend on your traffic mix, and a vendor quoting one for your account invented it.
- This is a cost-control product. We make no claims about revenue, returns or profit — not now, not later.
- There is no multi-tenant gateway you can route production traffic through today.
- Per-team and per-project attribution does not exist yet; the cap that exists is global to one router.
- Cheaper models save money and can cost quality. We make the trade-off visible, not automatic.
Questions people actually ask
Can we route our traffic through Tvijo today?
No. Portkey and LiteLLM ship that product; we do not. What we have is a router with a working cap and per-call cost records that we run ourselves, and a pilot conversation about applying the same pattern to your lanes.
Does adding a gateway slow requests down?
Any hop adds latency; the honest question is how much relative to model time and whether the control is worth it for that lane. Some lanes should not be governed at all, and we will say so.
Can we keep using our own provider keys?
BYOK is the intended design, and there is no key vault in the product today. Right now your provider relationship and invoice simply stay where they are.
What if our spend is small?
Then attribution matters more than caps, and you may not need us yet. We would rather say that than sell you a plan.
Find out what attribution would actually show
Describe the setup and we will tell you what a per-call record answers for you — and whether the change is worth making at your size.
Prefer to check us first? The record structure is written out field by field on the evidence page, and the external check shows its full report without an email.
Also useful
Published plans, so a pilot conversation starts from a number.
How Tvijo comparesPortkey, LiteLLM, Langfuse, LangSmith, Credo AI, n8n — what each does better.
Evidence pack structureField by field: what a record can hold, what exists today, what is design.
Book a walkthroughA person walks through what is built and what is not.