Claude¶

Claude AI content generation via Anthropic Messages API. Installing and configuring this extension registers Claude as a selectable AI provider platform-wide: admin content generation and prompt templates, AI-assisted SEO (meta and image alt text), the cart's AI order-list parser, and J'AI Chat replies all resolve their provider from the same registry. Claude supports both text and vision (image) requests.

What It Provides¶

• AI completion strategy — an AICompletionStrategy implementation registered under the name claude, calling https://api.anthropic.com/v1/messages with x-api-key authentication and a pinned anthropic-version: 2023-06-01 header.
• Vision support — supports_vision is True; multimodal requests send the image as a base64 image content block ahead of the text prompt.
• Connection testing — a build_test_strategy hook consumed by the testAiProviderConnection GraphQL mutation, so the admin can verify a key without persisting it.
• Hot reload — a hot_reload(config) hook: saving new config in the admin re-registers the strategy immediately, without an API restart.
• Settings deep link — the manifest points the Extensions list's Configure action at /settings/ai (AI Providers).

Once registered, the provider appears in the aiProviders GraphQL query (with a live health check) and is eligible for generateAiContent, SEO meta/alt-text generation, cartAiParseToLookupItems, and J'AI Chat — see Building Extensions for the strategy registry.

Setup¶

1. Enable the extension for a channel under Settings > Extensions.
2. Provide credentials — the install hint reads: "Set ANTHROPIC_API_KEY env var or configure via the extension settings." The config-form api_key takes precedence; the ANTHROPIC_API_KEY environment variable is the fallback when no key is saved.
3. Optionally choose a Model and Max Tokens default in the config form.
4. Save. The strategy hot-reloads from the decrypted config immediately.
5. Verify on the AI Providers page (Settings > AI) — the connection test makes a minimal one-token call against the configured model.

Configuration Reference¶

password-type fields are Fernet-encrypted at rest and masked (***) when read back through the admin API.

Permissions¶

The manifest declares no permissions.

Operational Notes¶

• No key, no provider — if neither the config api_key nor ANTHROPIC_API_KEY is set, the strategy is not registered at all (a warning is logged); Claude then does not appear in aiProviders and is skipped by auto provider selection. Saving a key later registers it without a restart.
• Image formats are whitelisted — vision requests accept only image/jpeg, image/png, image/gif, and image/webp (with image/jpg normalized to image/jpeg). Anything else — notably iPhone HEIC uploads — is rejected by the extension with a descriptive error before the Anthropic API is called; such images must be converted client-side before upload.
• Upstream errors are surfaced — on a 4xx/5xx vision response, the extension raises the Anthropic error body (e.g. "Could not process image") instead of a generic HTTP error, so admin-facing error details carry the real reason.
• Health check is a real call — health_check posts a one-token "ping" message to the configured model with a 10-second timeout, so a failing check can mean a bad key or an invalid model name.
• Timeouts — 60 s for text generation, 120 s for vision requests, 10 s for health checks.
• Masked-key test fallback — if the admin posts the masked *** value to the connection test, core falls back to the stored, decrypted key for that provider.
• Multi-channel config — the strategy is registered once per process; at boot the extension loads the first enabled channel row whose config contains an api_key.