Limits and provider keys

Quotas, per environment#

An org has one set of limits in production and another in the sandbox: minutes, messages, LLM tokens, calls at once, agents, remembered facts, knowledge chunks, bought numbers, seats, hosted apps, evals and simulations; and a budget in dollars, which both environments spend from. A limit nobody set is no limit; zero is a limit that refuses everything, which is how a plan leaves a feature out. The set is replaced whole, so a limit left out stops being one.

What an org used is counted from its calls' summaries and scores in that environment, every time a call, a turn, a judge or a simulated caller asks to start. The database keeps the count, one row per org, environment and calendar month (UTC) in usage_totals: a summary adds to it in the transaction that writes it, an erasure takes its calls out, and admission reads the month it is in, so minutes, messages and tokens start again on the first of each month (UTC), and a call is counted in the month its summary was written. Nothing is kept in memory, so a restart counts what the database holds; pinecall-runtime usage rebuild folds the table again from the summaries in the log. What the sandbox spent never closes production, except through the budget: dollars are the same in both environments.

• A call is refused past concurrent_calls, minutes, messages or llm_tokens. The calls at once are the platform's: every gateway counts its own and says them each second, so two gateways opening at the same moment may go past the limit by what they opened that second. One that is let in is told how many seconds are left of the minutes, and ends there.
• A budget (budget_usd, whole dollars a calendar month) refuses a new call once what the org's calls in both environments cost this month, as the summaries priced them, reaches it (the org has spent 10.5 of its 10 USD budget this month). A call already running is never cut for it, and written turns are not held to it.
• A written turn is held to messages and llm_tokens on every turn, with what the open conversation has spent so far added to the org's totals.
• Agents, numbers, facts, seats and chunks are refused at their limit; a push that would pass knowledge_chunks is refused whole.
• Hosted apps (hosted_apps) are counted when an app's first release is uploaded (Hosted apps); a later release of an app the org already has is never refused for it.
• Evals (evals) are the judges that answered a call (held, broken or classified) on the platform's key this month, a call's score and a judge tried on calls alike; a judge on the org's own key is never one. Past the limit a call still seals: the judges left are skipped, saying the month's evals are used up, and credits.exhausted lands once for that call.
• Simulations (simulations) are the calls a simulated caller played this month: a persona's, a spoken golden's. One past the limit is refused before it opens, and a run that reaches it stops where it is. A simulated call never counts minutes, and neither does a call an eval run opened.

A refusal is a 429 whose sentence names the quota, what was used and the limit (the org has used 30 of its 30 minutes in the sandbox), and credits.exhausted lands in the agent's log with the same numbers.

What a new org is given#

What an org gets when it is made is the platform's admission setting, edited from the console: the limits in each environment for a person's first org, and, when the platform gives one trial per person, the limits for any later org of the same address. A deployment with no such setting gives a new org no limits, which is what a self-hosted platform runs.

{
  "first": {
    "sandbox":    {"minutes": 30, "messages": 300, "llm_tokens": 2000000, "concurrent_calls": 1,
                   "evals": 500, "simulations": 10,
                   "lends": ["deepgram", "cartesia", "anthropic/claude-haiku-5-5"]},
    "production": {"minutes": 0, "messages": 0, "llm_tokens": 0, "evals": 0, "simulations": 0,
                   "lends": []}
  },
  "later": {
    "sandbox":    {"minutes": 0, "messages": 0, "llm_tokens": 0, "lends": []},
    "production": {"minutes": 0, "messages": 0, "llm_tokens": 0, "lends": []}
  }
}

Whose vendor keys a call runs on#

An org's own key for a vendor runs any model of that vendor, whatever the platform offers:

pinecall providers add deepgram        # the org's own key, sealed; every Deepgram model is theirs to run

Where an org brought no key, its calls run on the platform's, and only on what the org's lends says in that environment: null lends every key the platform holds, an empty list lends none, and otherwise each entry is a vendor (deepgram) or a vendor and a model prefix (anthropic/claude-haiku-5-5). A vendor nobody holds a key for is refused before the call opens.

A stage on a lent key runs as the operator configured it: the plugin's class and options are the providers row's alone. An agent that names its own (builds, options, from its class or its settings) runs that stage only on the org's own key, and is refused on a lent one — when the class registers, when the settings are set, and when a call's stages are handed out — naming the vendor and pinecall providers add. An option can point a plugin at another server; the platform's key goes to the vendor and nowhere else.

The simulated caller (its model and its voice) always runs on the platform's keys. The judge model does too, the providers row's judge, unless the org names its own (PUT /v1/org/judging) or the agent does (its judge): a model on a key of the org's own, a local one among them, is the org's to pay and its evals are never billed; one on a key the platform lends is lent as any stage is, by lends. A judge model the agent or the org names is tried when it is saved, on the org's keys in every environment it would judge in, so a model nobody can key is refused there and never fails a hang-up. Neither the judge nor the caller spends the org's minutes or tokens; the org pays for them as evals (one judge that answered one call on the platform's key: held, broken or classified, never na) and simulations (one call a simulated caller played), which call.score, evals.spent and call.summary carry and the usage feed counts (Charging for it §3), each held to its quota above.