Avunu AI addons
Jump to
↵select↑↓navigateescclose

MuK AI - Cloudflare AI Gateway Provider

On this page

Adds Cloudflare Workers AI as a selectable provider in MuK AI > Configuration > Providers, optionally routed through a Cloudflare AI Gateway so usage is tracked/billed under a named gateway.

Why

Cloudflare Workers AI advertises "OpenAI compatibility," but that only covers the older Chat Completions API (/v1/chat/completions) - MuK AI's own OpenAIProvider targets OpenAI's newer Responses API (/v1/responses) exclusively, which Cloudflare doesn't implement at all. Subclassing that adapter therefore doesn't work: it fails validation on Cloudflare's side the moment a real conversation (not just a model listing) is attempted. This module implements its own provider against Chat Completions directly instead - the same "not the Responses API" translation problem MuK AI's own Anthropic provider already solves for a different vendor, and this module's structural template for it.

On top of that, Cloudflare's account id is all that's actually needed to use it: the endpoint URL and the account's available models can both be derived from it, rather than hand-typed.

What it does

  • Registers a new cloudflare provider, implementing Chat Completions' request/response/streaming shapes directly (tool-calling included) against a plain ProviderBase subclass, that takes an Account ID instead of a URL, and builds https://api.cloudflare.com/client/v4/accounts/<id>/ai/v1 from it - the Endpoint field below still shows the result, it's just no longer something you type.

  • Replaces the generic API Key field (hidden for this provider) with API Token - the Cloudflare API token itself, sent as Authorization: Bearer <token>.

  • Gateway ID, if set, is sent as the cf-aig-gateway-id header, routing the request through that AI Gateway for usage tracking - no URL change needed.

  • A Fetch Models button (next to Test Connection) creates/updates a muk_ai.model record for each chat or image model available (everything else - embeddings, ASR, TTS, etc. - is skipped, since muk_ai only catalogs chat and image models). Running it again updates existing entries rather than duplicating them.

    • If Gateway ID is set and that gateway has a spend-limit rule restricting Workers AI to specific models, only those models are fetched - matching what the gateway will actually serve.

    • Otherwise (no Gateway ID, or a gateway with no such restriction), every model in the account's public catalog is fetched.

What it does not do

Cloudflare's model search API doesn't return pricing, so every fetched model comes in with its input/output rate at 0 and a note on the record saying so. Review and correct pricing manually (developers.cloudflare.com/workers-ai/platform/pricing/) before relying on MuK AI's per-session cost tracking for any Cloudflare model.

Configuration

  1. MuK AI > Configuration > Providers, create a provider with Provider = Cloudflare Workers AI.

  2. Account ID = your Cloudflare account id (Account Home, right-hand sidebar).

  3. API Token = a Cloudflare API token with Workers AI permissions.

  4. Gateway ID = the gateway's ID from Account Home > AI Gateway, if you want usage tracked under it. Leave empty to call Workers AI directly.

  5. Test Connection, then Fetch Models to populate the catalog - then open each imported model and fill in its real pricing.