Getting started

Siftline asks a model a fixed set of questions about each piece of text and gives you the answers as plain values: a label, a yes/no, a level. You write the questions once, measure how well they are answered on examples you label, then run them on real text and route each answer to a webhook, a Slack channel, or nowhere. There is no generated prose anywhere in the loop.

The toolkit is three npm packages, all ESM-only, all on Node 22.14 or newer.

Install

The guide starts with the CLI, which needs no code:

npm install --save-dev @siftline/cli

Set your TypeSafe API key in the environment. The CLI reads it from there and nowhere else:

export TYPESAFE_API_KEY=...
npx siftline --version

Add @siftline/core when you want to do the same from TypeScript, and @siftline/actions when a Decision has to go somewhere:

npm install @siftline/core @siftline/actions @typesafe-ai/sdk

@typesafe-ai/sdk is the client the Engine talks through. The CLI bundles it; from code you construct it, as The same from code shows.

Where to go next

• Concepts defines the eleven words the rest of the site uses.
• The guide builds one support inbox end to end, in the order you would work: write the questions, measure them, run them, route, send.
• Packages lists what each package exports and how it fails.
• The API reference is generated from the source and always matches the published types.