@a11yci/llm

Optional, bring-your-own-key Claude adapter for a11y-ci. It adds two advisory capabilities on top of the deterministic engine:

• Fix suggestions — a concrete, minimal fix for each detected issue.
• Semantic review — the issues automated checkers can't judge: unhelpful alt text, non-descriptive link text, label-in-name mismatches, sensory-only instructions.

Both are advisory. Suggestions are attached to issues for humans to review; semantic findings are emitted at "manual" severity, so they can never flip the deterministic pass/fail gate.

import { audit } from '@a11yci/core';
import { createAnthropicAdapter } from '@a11yci/llm';

const llm = createAnthropicAdapter(); // reads ANTHROPIC_API_KEY; returns null if absent
const result = await audit('https://example.com', { llm: llm ?? undefined });

if (llm?.suggestFix) {
  const fix = await llm.suggestFix(result.issues[0]);
}

Set ANTHROPIC_API_KEY in your environment. The key never leaves your machine / CI runner; this package calls the Anthropic API directly and ships no telemetry.

Source-available under the Elastic License 2.0 (see LICENSE).