Limits

Two model providers, a fallback and a reserve for each agent role. OpenAI writes and judges; when it's down or out of credit, Nemotron takes over, slower on NVIDIA's free tier (about 40 requests a minute for each model, split between the web app and the worker). If every model of a role is busy, the answer says so and asks you to try again. Sample mode needs no key at all.
Live use needs Zetaris. Nothing is marked verified without it, by design.
"Supported" means the cited passage states it, not that the claim is true in the world. A wrong source, quoted correctly, still passes.
Sample mode replays recorded questions only. Any other question is refused, not invented.
Fair-use caps, no paywall: 300 answers, 20 deep searches and 300 files a day per account; guests get less (60, 3, 40). Settings → Usage shows them.
Connectors use Composio-managed Google auth, whose consent screen asks for wider Gmail scopes than Konkyo uses (it only reads). A read-only setup needs your own Google OAuth client.
Scanned PDFs are read with OCRmyPDF, in English; handwriting and complex tables can come out imperfect.
Check my draft reads up to 20 claims from the first 30,000 characters of a note, and a published article shows what that check found; claims it didn't support still publish, counted.
Deep search reads only open-access papers it can download; paywalled publishers are skipped.
Email goes through Resend's free plan: 100 emails a day, sent and received together. So sign-in codes stop at 50 a day (8 per network address), invites by email at 30 (10 per person), and Google sign-in is always there when codes run out.