Skip to main content
Every turn is metered, and every organization has a budget. Karta tracks what your agents cost, lets you set caps that stop spend before it happens, and returns 402 Payment Required when a cap is hit - no surprise bills.

​
What’s metered

Each agent turn records input tokens and output tokens, attributed to the organization and, when available, the API key, karta, and verified acting user involved in the turn. Total tokens and cost are derived - cost from the per-model rates on the rate card. Metering is best-effort and never blocks your agent’s response: usage is buffered and flushed in the background, and successful retries are de-duplicated so nothing is double-counted. Under a management-service outage or extreme load, a small fraction of events may be dropped.

​
Budgets and cost thresholds

An organization budget can cap input tokens, output tokens, total tokens, and cost. API keys can also carry sub-limits for input tokens, output tokens, and cost, so one integration cannot burn the whole budget. Enforcement gates on two things:
  1. Subscription status - an inactive subscription blocks consumption.
  2. Caps - any configured limit being exceeded blocks consumption.
When a cap is crossed, the next request returns:
Enforcement happens before new work starts. Usage metering is asynchronous, so a turn already in flight can finish; the next gated request sees the updated budget state.

​
Karta-grain caps

Beyond the organization budget and API-key sub-limits, you can cap spend at the same grains your product uses to create kartas:
  • Per user - a hard monthly cap on one verified user, across every per-user karta they use. Set it from the End-users page.
  • Per karta - a cap on one durable karta, such as a virtual employee or backend worker, summed across everyone or everything using it. Set it from the Kartas page.
  • Per seat - a cap on one verified user’s spend inside a shared virtual employee karta, also set from the Kartas page.
All of them are hard caps: when one is reached, the next request at that grain is refused before it runs - a 402 on the API, or a graceful limit_reached state in the embedded widget. Caps you have not set never block anything. Every cap in effect - organization, API key, per-user, per-karta, per-seat, and the per-embed-key anonymous ceiling - is listed together on the Limits page (and on Usage), each linking to where you set it. Start there to see what’s capped and what is not.

​
Credits and the prepaid gate

On managed plans, usage draws down a prepaid credit balance - there is no metered overage and no invoice. When the balance runs out, requests return 402 with reason credits_exhausted until you top up. Bringing your own model key (BYOK) means you pay your provider directly for model tokens, but Karta’s platform fee still draws down the same prepaid balance - so the credits_exhausted gate applies there too. See Billing & credits for plans and packs.

​
Threshold alerts

Register a webhook to be notified as you approach or cross a budget threshold (and on billing events). Use it to raise a cap, alert a team, or throttle traffic before you hit a hard stop.

​
Reading usage & cost

Use these for cost reporting and capacity planning - track period-over-period spend, attribute cost to individual keys, and forecast against your caps.

API keys

Per-key sub-limits and scopes.

Webhooks

Get notified on budget thresholds and billing events.