sayintel
Docs

Models.

sayintel routes by task tier rather than using one model for everything. Judgement-heavy writing gets the strongest model; the high-volume classification around it gets a fast one. The split is why a campaign of a few hundred people stays affordable without the openers reading like a template.

The routing

plumbing   fast model   prequalify · qualify · research · signal picking ·
                        quality checks · scraping vision · classification
drafting   Claude       the three LinkedIn messages, rewrites, email sequences
fallback   fast model   used only if the drafting call fails, so writing never stops

Both tiers are configurable per deployment, so the exact model names move as better ones ship. What does not move is the shape: one tier for volume, one for voice, and a fallback so a provider outage degrades quality instead of stopping the pipeline.

Bring your own key

Drafting runs on sayintel's own Claude access on every plan, including free. On the higher plans an account can add its own Anthropic key instead, for isolation and its own budget; key resolution is your key first, then the workspace key, then the fallback model.

Streaming

Written output streams token by token, so a rewrite of eighteen openers shows progress per person rather than a spinner. Long pipeline stages report as they advance, through the same event stream an agent can subscribe to — see webhooks and events.

Cost and caps

Every paid call records its own cost against the account. A margin guard stops paid work rather than overspending, and spending capabilities expose a cost_preview plus dry_run so an agent can size the bill before committing. Details in the agent quickstart.

Quality rules the models cannot break

  • An opener is anchored on a verified extracted signal, or it says something honestly general. There is no third option, and no invented news.
  • Drafts pass a hard-fail gate before they reach a human: word caps, banned phrases, fabricated proper nouns, recycled proof points. Every code is published in errors and QA codes.
  • Nothing sends without a human tap, whichever model wrote it.