In the console (hosted plans, or self-hosted Enterprise): Settings, then “AI cost”. The cap is set on the same screen, by an owner or an administrator.
On the command line (free self-host): kastel cost show for this month's spend, kastel cost estimate to size a corpus before reading it.
Who pays for what
Your Kastel works with your company's AI keys. You hold the account at the AI provider, they invoice you, and Kastel never resells inference - neither hosted nor self-hosted. That is not a billing option, it is a product rule: nobody should make money by standing between you and the model you chose.
The practical consequence: you can change AI provider without changing Kastel, and your AI bill stays readable at your provider, line by line.
The monthly cap
You can set a monthly spending cap, and an alert threshold as a percentage of it. The cap is optional: not setting one is a legitimate choice, and your Kastel says so rather than inventing one.
When the month's cap is reached, new AI operations are paused. The screen says it in those terms: it is the cap you set, not a Kastel limit. You raise it whenever you want; changing a cap is a traced action.
Search and indexing: the calls that find the right information inside your Kastel. Blocking them on a cap would make your Kastel mute at the worst moment, for by far the smallest spend in the product. Lesson compilation: it runs in bounded batches and stops on its own budget. Both appear in your spend statement: they are exceptions to the block, never hidden spending.
Nothing starts without a quote
Any operation about to spend seriously shows its estimated cost before offering the button: reading a new scope of documents, deepening a department, running a contradiction check, launching the statement reading, measuring the Context Score.
The quote is always a range in euros, together with what is left of your monthly cap. An upper bound alone would turn a cautious estimate into a price; a lower bound alone would lie in the expensive direction. If the upper bound does not fit in what is left, nothing starts and the page says so before offering anything.
The quote is deliberately cautious: what has already been read and has not changed is never re-read or paid for twice, but the quote does not deduct that saving in advance. You therefore end up under the quote more often than over it.
Where to read the spend
The “AI cost” screen shows what your Kastel cost this month, and where the money went: by operation (setup, ingestion, consultations, agents, self-improvement), by model, and by use. It also shows the reading side: how many times your connected AIs read your Kastel this month. That part costs almost nothing; it is displayed for transparency.
On a Kastel you host, kastel cost show gives the same statement grouped by use, and kastel cost estimate sizes a corpus before reading it - useful to know what a new department would cost before opening it.
The Context Score answers the neighbouring question - “is my Kastel any good” - with its own spending guard-rails. It is described with day-to-day upkeep.
Read the Keeping your Kastel in good shape pageAn estimate, never an invoice
Every amount Kastel displays is an estimate, computed from the volume processed and an indicative price table. Your AI provider's invoice is the truth, always, and Kastel does not claim to replace it.
These figures serve one purpose: sparing you a surprise. They are never billed, never sold, never transmitted. Nor do they measure your usage for commercial purposes - that is written into the product, not only into a contract.
What Kastel itself costs is a different question, and it has its own page.
See pricing