FAQ

Yes — to read your description and pre-fill the wizard fields. No — for the cost math. The forecast is pure arithmetic on counted tokens times published rates. A page like /how-it-works walks through every step. Inferred values show a “PitCrew suggested” pill so you can override anything that looks wrong.

Every dollar figure has a low/mid/high confidence band. The band widens when an input was inferred (vs typed) or when the user said “I don’t know.” Token counts come from real BPE tokenizers (not the chars/4 heuristic) where the provider exposes one, falling back to a stress-tested estimator otherwise. Pre-deploy guesses on call volume miss by ±50% on average — the band reflects that.

It’s stored attached to your audit, behind your account. The sign-in cookie scopes access — only you (or a recipient of a URL with your session cookie) sees your reports. Demo reports at /demos are the only public ones, and we hand-curated those. Settings → Delete my data permanently removes your account and every audit attached to it.

We use list pricing from each provider’s public rate card. If you’ve negotiated a 30%-off enterprise contract, your real numbers will be 30% lower across the board — but the relative ranking of optimizations (cache vs. batch vs. model swap) won’t change. The confidence band on the savings number is much wider than typical contract discounts.

We cite the cascade row by row. Each recommendation shows what it saves and what it might break — “Switch to Haiku 4.5 from Sonnet 4.6” carries a flag about A/B-testing quality before fully switching. The engine won’t recommend a move it can’t justify; on some agents (frontier-quality coding work, for example), the demo at /demos honestly says “you’re already optimized” and suggests no change.

Anthropic, OpenAI, Google, xAI, Meta, Mistral, DeepSeek, MiniMax, Moonshot for the text lane; plus generation rows for video (Veo, Runway), image (Flux, DALL-E), TTS (ElevenLabs, Cartesia), and music (Suno, Udio). The full list is browsable at /models. Each row shows when the pricing was last verified.

No. PitCrew is a pre-deploy forecasting tool — it doesn’t take provider API keys, doesn’t ingest live usage data, doesn’t analyze deployed agents. The forecast is built from your description plus published rate cards, full stop. If a future version adds post-deploy analysis it’ll be opt-in and called out separately.

Pick “Custom” on Step 0 and fill in the fields by hand. The parser can also handle freeform descriptions of unusual setups — if the inferred archetype is wrong, click the suggested pill on Step 0 and pick a closer one. Multimodal generation (video/image/TTS) gets its own Step 6 when you select a generation archetype.