Skip to main content
MCP spend comes out of your Developer Balance — shared with the REST API, separate from consumer subscription credits. Top up at Developers → Billing, or let the agent do it: add_credits returns a Stripe checkout_url right in the conversation.
Browsing is free — connect and ask questions without a balance. Your first paid action can unlock $0.50 starter balance with a quick phone verify, and NinjaChat subscribers get $1.00 of MCP allowance every month. The allowance is granted lazily — the first time a charge would otherwise fail, or when get_account_balance is called — to accounts on a paid plan (a mobile plan must still be active).

​
Prices

Model prices are the providers’ own rates, passed through unchanged. NinjaChat’s margin is a single 8% service fee added when you fund credits — $25 of credit costs $27 at checkout — never a markup hidden inside a model’s price. Multiply by n for multi-output calls. Every video model has its own duration curve. Per-second models (seedance-lite, seedance-pro, seedance-2, runway-gen4.5) bill each second up to their cap; bucketed models (Veo, Kling) bill the nearest supported length — 7 seconds on veo-3.1-fast costs the 6-second price, 8 seconds on kling-video the 10-second price. Any duration from 4 to 15 is accepted, but seconds past a model’s cap are billed at the cap. estimate_cost is free and returns the exact price before anything runs — pass duration for video.

​
Money is safe by design

Hard budgets

max_spend_cents (1–2000) on any call, total_budget_cents (5–100) on bakeoffs — tools return budget_too_low instead of exceeding them, before any money moves.

Automatic refunds

Image and bakeoff failures refund inline. A video job that fails after being charged is refunded when polled — and a server-side sweeper checks every minute even if nobody polls. You never pay for output you didn’t get.

Safe retries

Pass an idempotency_key — the same key + body never double-bills.

Monthly limit

Set an account cap in Developers → Billing — the OAuth consent screen also sets one the first time an agent connects. It’s reserved before deduction — an over-limit call never touches your balance.

​
What a charge actually does

Every billed call runs the same sequence: idempotency claim → max_spend_cents gate → account monthly-limit reserve → per-key budget reserve → deduct → generate → settle against the model that actually ran → store the response for replay. If generation fails, the refund lands before the error tells you that you weren’t charged, and every reservation is released in reverse order.
  • Settlement is exact. auto may reroute to another model when a provider fails; the charge is adjusted to that model’s price, and cost.charged_cents in the response is what you actually paid.
  • Two spend caps. spend_limit_exceeded is the account cap from Developers → Billing. key_budget_exceeded is a per-key monthly budget (Developers → Keys → Manage) that applies to API-key callers and covers everything the key does, REST and MCP alike. Both errors return the limit and this month’s spend; neither touches your balance.
  • Empty balance. insufficient_credits returns balance_cents, estimated_cost_cents and top_up_url. The fix is add_credits, then retry.

​
Idempotency

  • Keys are scoped to your account and bound to a hash of the billable inputs (prompt, model, count, aspect ratio, references, duration…). They live for 1 hour.
  • Same key, same body → the stored response is replayed, images re-rendered inline, nothing charged.
  • Same key, different body → idempotency_conflict. Use a fresh key for new content.
  • Same key while the first attempt is still running → generation_pending with retry_after_seconds: 10.
  • A failed generation releases the key, so a retry with the same key runs (and is charged) once more.
  • If the idempotency store is unreachable the call fails closed with idempotency_unavailable — no work is started, nothing is charged; retry in a few seconds with the same key.

​
Refunds

Video jobs carry the exact settled charge, so an async refund always returns what was actually paid. get_generation_status on a refunded job keeps reporting the failure — a refunded job never later serves a video.

​
Topping up

  • add_credits mints a Stripe Checkout link for a $25 (default), $100 or $250 pack — the agent pastes checkout_url, the user pays, the agent retries. The Billing page is the fallback and also takes custom whole-dollar amounts.
  • Auto-reload is opt-in and only ever switched on from the Billing page — buying a pack never enables it. Once on, it defaults to adding $100 when the balance falls below $10 (both adjustable).
  • Every purchase carries the 8% service fee: add_credits with amount_usd: 25 charges $27 and credits $25.

​
Rate limits

Counted per API key — or, for OAuth clients, per client + user — so one busy agent can’t starve another. Per minute unless noted: Limited calls return rate_limited with retry_after_seconds — agents wait and retry.

​
Agents

Every MCP client — Cursor, Claude Code, Codex, VS Code, claude.ai, ChatGPT, Hermes, OpenClaw, and anything else on the same server — draws from the same Developer Balance and the same monthly limit. Rate limits are counted per connected client. Review or disconnect anytime in Developers → Agents.