@richkitjs/extension-ai

AI authoring for RichKit. Streams a model's output into the document, and — when @richkitjs/extension-track-changes is installed — lands every edit as a suggestion the user can accept or reject.

The extension never talks to a provider itself. You hand it a complete transport, which is the only vendor-specific seam:

• @richkitjs/ai-openai — OpenAI chat completions
• @richkitjs/ai-anthropic — Anthropic Messages API

Install

pnpm add @richkitjs/extension-ai @richkitjs/ai-openai

Usage

import { Editor } from '@richkitjs/core'
import { StarterKit } from '@richkitjs/starter-kit'
import { AI } from '@richkitjs/extension-ai'
import { openaiComplete } from '@richkitjs/ai-openai'

const editor = new Editor({
  element: document.querySelector('#editor')!,
  extensions: [
    StarterKit,
    AI.configure({
      // Your own route, holding the API key server-side.
      complete: openaiComplete({ endpoint: '/api/ai' }),
      attributeAs: 'AI Assistant',
      track: true,
    }),
  ],
})

// Rewrite the selection.
editor.commands.aiPrompt({ prompt: 'Make this more concise' })

Options

| Option | Type | Default | Description | | ------------- | -------------------- | ---------------- | ------------------------------------------------------------------------------- | | complete | AIComplete \| null | null | Transport. Commands no-op while this is null. | | attributeAs | string | 'AI Assistant' | Author recorded on the suggestions it produces. | | track | boolean | true | Route output through track-changes. Inert when that extension is not installed. |

Commands

| Command | Description | | ------------------------------ | ------------------------------------------------------------------------------- | | aiPrompt({ prompt, range? }) | Run against range, or the current selection. No-ops while a run is streaming. | | aiCancel() | Abort the in-flight run. The transport's signal fires. | | aiRetry() | Drop the last attempt and re-run the same prompt against the same target. | | aiAccept() | Keep the AI text: drop what it replaced, strip the suggestion marks. | | aiReject() | Discard the AI text and restore what it replaced. |

With track: true, a rewrite marks the original text deleted rather than removing it, so aiReject can restore it. Without tracking there is nothing to fall back to, and the original is deleted outright.

Plugin state

import { getAIState, isAIStreaming } from '@richkitjs/extension-ai'

const state = getAIState(editor.state)
// { status: 'idle' | 'streaming' | 'error', prompt, error, range, replaced }

if (isAIStreaming(editor.state)) {
  // show a stop button
}

Transport failures — non-2xx responses, stream errors, refusals — surface as status: 'error' with the thrown message in error.

Writing a transport

Any async generator of text chunks works:

import type { AIComplete } from '@richkitjs/extension-ai'

const complete: AIComplete = async function* (req, { signal }) {
  // req: { prompt, selection, documentText }
  const res = await fetch('/api/ai', {
    method: 'POST',
    body: JSON.stringify(req),
    signal,
  })
  for await (const chunk of readChunks(res.body!)) {
    if (signal.aborted) return
    yield chunk
  }
}

selection is the text being rewritten, or '' when inserting at a cursor. Transports must honour signal so aiCancel actually stops the request.

License

Part of RichKit Pro, licensed under the RichKit Pro License Agreement. It runs without a key on localhost and other development hosts. Production use needs a licence key — plans start at $99 a year, see the pricing page.

import { setLicenseKey } from '@richkitjs/license'

setLicenseKey('YOUR-LICENCE-KEY')

The key is checked locally; nothing is sent anywhere. Without one, a small "unlicensed" badge shows on production sites — the editor never stops working.