Each department opened its own
Marketing has a subscription, the dev has an API key, someone else paid for another one on a card. Nobody can say how much the company spends on AI per month.
All services
The gate · Router
Every AI in your company enters through the same gate. And it picks the cheapest one that solves it.
One key, the entire catalog of models. Each request is evaluated before it goes out: can it be answered from cache, with a free model, or with the subscription the company already pays for? The expensive API is the last resort, and the dashboard shows how much that kept from leaving the cash register.
ONE KEY FOROpenAIAnthropicGoogle+8 providers
A REQUEST COMES IN
"Summarize this email in three lines."
Cachethe same question was already answered before
Free modeleleven providers at no cost per call
Subscription you already pay forthe account the company already has, used up to the limit
Paid APIlast resort, when the task really requires it
COST AVOIDED ON THIS REQUESTUS$ 0,004what it would have cost if it had gone straight to the paid API
Example of routing. On your dashboard the value is the real one, summed by period, and the local layer doesn't appear because it's the appliance owner's hardware, not the client's.
Nobody picks a model for every request. The system picks.
The AI bill doesn't blow up because of one expensive request. It blows up because a thousand cheap requests went to the most expensive model that was configured. The Router evaluates each one before it goes out, and uses the cheapest layer that can handle it.
What happens today
Three AI accounts, none of them knowing what the other spent.
The expensive model answers cheap questions
The configured key is the one that handles everything. Summarizing an email and building a strategy go through the same place, at the same price.
When the provider goes down, the operation goes down with it
One key, one provider, one point of failure. Its instability becomes your problem, and there's no backup plan configured.
What you start to see
The whole company's AI bill, on one screen.
Every call that passes through the gateway gets logged: who requested it, which layer resolved it, how much it cost, and how much it would have cost. It's what turns "we spend too much on AI" into a sentence with a number behind it.
EXAMPLE OF A SNAPSHOT
0%
Cost avoided0%
Availability0
Models available0
Departments usingCost avoided
How much the company avoided spending because the request was resolved on a cheaper layer. It's the metric that justifies the module on its own.
It's the number you bring to the cost meeting
Availability
If a provider goes down, traffic reroutes to the next one without anyone getting woken up. You don't inherit any vendor's downtime.
Their downtime stops being yours
Models available
The whole catalog behind a single key, with the same call format. Switching models stops being a migration project.
Testing a new model costs one line
Departments using
Who consumed what, by department and by project, with a cap per key, so no one drains the balance by accident.
You can grant access without giving a blank check
Under the hood
How it decides, in four questions.
The decision is made before the call goes out, and the order never changes: it starts with what's free and only moves to paid when the options above can't cover it.
Has this already been asked?
An identical request with a stored answer comes back from cache. It costs zero and arrives instantly: it's the layer that saves the most on repetitive operations.
resolves without even leaving your infrastructure
Can a free model handle it?
Most routine work doesn't need the top tier model: summarizing, classifying, extracting. If a free one delivers, it's the one that answers.
the biggest share of volume lands here
Does it fit the subscription you already pay for?
The company already has AI accounts. They go into the pool and are used up to the limit before any new cent is spent.
you start using what you were already paying for
Only then: paid API.
When the task really requires the expensive model, it's called, and the dashboard records that the spend was necessary there, not an oversight.
the expensive becomes a justified exception
To clear it up
Four comparisons that settle this.
A direct key works, a third-party gateway works, and having a dev handle it works. The point is how much each one costs as volume grows.
Direct key with the providerThrough the single gateway
Cost per requestAlways the price of the configured model The cheapest one that handles that request
If the provider goes downYour operation goes down with it Reroutes to the next one automatically
Switching modelsChange code and test again One line of configuration
From zero to the first snapshot
Three steps. Your first call in minutes.
011 MINUTE
Create the key
One Rovemark key replaces the provider keys. You set its spending cap when you create it.
025 MINUTES
Change the address
The gateway accepts the format your code already uses. In practice you change the base URL and the key: the rest of the code stays the same.
03ON THE SAME DAY
See the avoided cost
From the first call, the dashboard shows which layer handled each request and how much was kept from leaving.
Inside the company
The same door, three different kinds of relief.
You get a number for "how much we spend on AI," plus a cap per key, so spend doesn't depend on good judgment.
Stop maintaining an integration per provider and waking up when one goes down. One endpoint, and failover is automatic.
They don't choose a model, don't think about price, don't ask anyone for a key. They ask for what they need and get it.
What comes out of this, concretely
01
The monthly statement
Spend and cost avoided by area, project and period.
02
Keys with a cap
One per person or per project, revocable, with its own limit.
03
History of every call
Which layer resolved it, what it cost, what it would have cost.
Where the alert arrives
When a key goes over its cap or a provider starts failing, the alert reaches the team wherever they already are.
SlackEmailWhatsApp
View Connect How much it costs
You pay for the model. The gateway doesn't charge a toll.
The Router doesn't add a margin per call: the credit consumed is the cost of the model that answered. When the answer comes from cache, the free tier or your own subscription, there is no credit at all.
Resolved in cacheA question that already had a saved answer.
0 credits
Resolved on the free tier or your own subscriptionMost routine volume falls here.
0 credits
Resolved on the paid APIOnly when the task requires the expensive model.
custo credits
The free plan already connects the Router to models with no cost per call. You can run the entire routine operation without spending a single credit, and see on the dashboard what it would have cost elsewhere.
See all plansBefore you ask
What people usually want to know.
Do I need to rewrite my code?
No. The gateway accepts the format most libraries already use, so in practice you just swap the base URL and the key. What used to call a provider directly now calls the gateway, and the routing happens on the other side.
What if the routing picks a worse model?
The choice is based on capability, not chance: each type of task has a quality floor, and a model only qualifies if it meets that floor. You can also pin a specific model to a route whenever you want: savings is the default, not an obligation.
Can I use my current subscriptions?
That's exactly the point. The accounts the company already pays for join the pool and get used up to their limit before any new spend: a lot of people discover here that they were paying twice for the same capacity.
What happens if the gateway goes down?
It degrades cleanly: it returns an explicit unavailability response instead of hanging, and your code handles it the same way it would handle any provider going down. Nothing sits there waiting.
Do you store the content of my requests?
What gets stored is yours, encrypted with your key, and exists for the cache and for the history you consult yourself. It comes out in full whenever you want, and it never becomes training data for anyone's model.
THE OTHER ORGANS
The gateway serves the other modules.
Every request from every module goes out through here, so the Router's savings show up on the whole system's bill, not just for whoever calls the API by hand.
THE OTHER ORGANS
No module works alone.
Each one solves a front, and all write to the same memory. See the others and how they fit with this one.