MIDNIGHT/ docs
PricingSign inStart free

Credits & limits

One credit is one message on the model serving today. Every model's cost is printed on its card before you pick it, so the number you see when choosing is the number you are charged.

What a message costs

ModelCredits per message
Gemini Flash1
One model is live today
Gemini Flash is the only model currently connected, so it is the only one the picker offers and the only cost there is to know. Any model added later appears in the picker with its own multiplier, and only once it can actually serve traffic — so you cannot select, or be charged for, a model that would fail to answer. The card in the Behavior tab is always the authority on what a model costs.

Two kinds of message are free:

  • A grounded fallback — the agent judged its retrieval too weak and said so instead of guessing.
  • Anything sent while a human has taken over the conversation.

Plans

PlanPriceCredits per monthAgents
Free$01001
Starter$19/mo2,0001
Growth$59/mo8,0005
Scale$199/mo30,000Unlimited, plus white label

Credits reset monthly. Beyond your allowance, overage is $15 per 1,000 credits, billed in whole blocks of 1,000. You get a warning in the dashboard once under 10% remains.

Credits are checked before a message and deducted after it, in the same transaction that stores the reply. A workspace that runs out mid-message can therefore overshoot by at most one message — never more, and never silently.

Rate limits

LimitWindowScope
20 messages5 minutesPer visitor session
60 messages1 hourPer IP address, per agent

Both apply to normal traffic without anyone noticing; they exist to stop a script. A visitor who trips one gets a rate_limited error, which is retryable.

The Retry-After header on a 429 carries the seconds until the window actually frees a slot, so it is safe to show a visitor. Note that a client which ignores it and keeps retrying records those attempts too, and can push its own wait out.

Other caps worth knowing

  • A message is at most 4,000 characters.
  • A widget session token lasts 30 minutes; the widget renews it transparently.
  • A website crawl covers up to 50 pages; auto-sync processes 10 sources per run.
  • Uploads are capped at 20 MB per file.
API reference
Endpoints, payloads, and error codes.
Sources & training
What the agent knows.