Questions, answered straight.

No. BurnCap is never a proxy — your model calls never pass through it. You send usage out-of-band through an OpenAI or Anthropic import, a CSV upload, or our SDK. Budget checks default to fail open if BurnCap is unreachable; fail-closed behavior is an explicit SDK choice.

Three ways, and you can mix them: pull daily totals from the OpenAI or Anthropic admin APIs, upload a usage CSV, or track per-request with the @burncap/sdk. Ingestion is idempotent by request id, so retries and re-uploads never double-count.

OpenAI and Anthropic have native daily imports. Any other provider — Gemini, OpenRouter, LiteLLM, Vercel AI Gateway, your own gateway — works through the SDK or a CSV export.

Yes. Pick your stack on the Agent prompts page and paste the ready-made prompt into Claude Code, Cursor, or another coding agent. It uses provider-aware SDK adapters where available, explicit mappings otherwise, and can add optional budget checks.

You can start with a CSV, connect a supported provider admin key, or add the SDK at your server-side model call sites. You can also load demo data during onboarding to explore the product before connecting anything.

No. BurnCap stores token counts, costs, and only the identifiers you choose to send — feature, customer_id, environment, request id. It never receives or stores your prompts or model completions. Ever.

Costs come from an effective-dated list-price catalog plus any pricing overrides you set for negotiated rates. Unknown models stay unpriced, uncovered meters are marked partial, forecasts report their evidence limits, and reconciliation compares tracked spend with the provider-billed ledger.

Yes. One click in Settings deletes all usage data for your workspace immediately. Provider keys are encrypted at rest and API keys are stored only as hashes.

BurnCap returns fresh budget state and an advisory action; your code decides what to do — block, warn, or drop to a cheaper model. BurnCap never blocks traffic itself, and the SDK defaults to fail open, so enforcement and outage behavior stay under your control.

Your app can reserve a conservative USD amount before an expensive model call. BurnCap checks and creates that short-lived hold atomically, counts active holds alongside tracked spend, and replaces the hold when your app reports the completed call. The hold expires if the call never happens; your app still makes the final enforcement decision.

A single 0-100 readiness number answering 'are we set up to catch a cost surprise?' It checks signals like whether you have a budget, whether your data is fresh, whether alerts can actually be delivered, and whether spend is tagged — then points you at what to fix.

Daily spend spikes versus your trailing average, runaway request loops per customer or agent session, expensive-model usage, dev/test spend, and unprofitable-customer margin alerts. Deliveries go to email, and to Slack on paid plans.

Each eligible account can receive one 30-day Starter grant after signup and, when required, email verification. No card or checkout is needed, nothing renews automatically, and no charge occurs at expiry. Unless you voluntarily upgrade, the workspace returns to Free limits with its existing data intact. Existing paid or previous promotion recipients are not eligible.

Yes — free forever, no credit card. The Free plan includes 10k ingested usage records per month, 7-day history, one project and one seat, budgets and caps, and spike and runaway-agent alerts by email.

Starter ($19/mo) adds daily auto-sync, Slack alerts, more usage records and history, cost insights with what-if model swaps, and a weekly burn report. Growth ($39/mo) adds margin alerts that flag unprofitable customers, a daily burn report, a full year of history, and more projects and seats.

Paid plans are billed monthly through Lemon Squeezy, our merchant of record. Cancel anytime from the in-app billing portal — you keep paid access until the end of your period, then revert to Free with your data intact.

Solo founders, AI wrapper products, agencies shipping AI tools, and small product teams of one to ten people — anyone spending real money on LLMs without a dedicated FinOps function.