One key per provider. Pick a provider/model per task. Every save runs a live probe; policy contracts prevent an AI role from gaining accounting authority.
API keys live here, once per provider. Slot routing below uses these by reference.
/v1/models, /v1/chat/completions). HTTPS only — no localhost or private/internal hosts.Every row shows the effective provider, model, endpoint, and key source resolved from live settings. Mismatched model/provider pairs are blocked rather than silently rerouted.
Each task has a versioned role, reasoning level, output contract, tool boundary, and hard output-token ceiling. Policies can narrow behavior only; they cannot grant posting, payment, tax-filing, or external-send authority.
Append-only versioning. Saving creates a new version and retires the prior. Click to expand.
Mistral OCR is a separate product from chat completions — it has its own client. Used by the document-ingest pipeline for PDFs and images.
Live token + cost telemetry from ai_call_log. 24h / 30d totals plus a per-job breakdown so a runaway cron is visible before it bills.
Per-job and global daily/monthly caps. When a cap is hit, the job logs status='skipped_cap' and skips its API call until the next window.
When ON, Plaid transactions matched to pre-staged drafts auto-post. Otherwise we link only.