pi-cloudflare-workers-ai

A standalone pi extension for Cloudflare Workers AI.

It wires Cloudflare's free-tier models into pi as a single cloudflare-workers-ai provider:

Capability	Transport	Status
Chat / reasoning (text)	OpenAI-compatible …/ai/v1/chat/completions	✅ delegated to pi's openai-completions core
Image generation (FLUX.1 Schnell)	…/ai/run/@cf/black-forest-labs/flux-1-schnell	✅ verified shape
Text-to-speech (Deepgram Aura 2)	…/ai/run/@cf/deepgram/aura-2-*	✅ verified shape
Speech-to-text (Whisper / Deepgram Nova-3)	…/ai/run/@cf/openai/whisper, …/ai/run/@cf/deepgram/nova-3	✅ verified transports

Split out of pi-cn-free-model-providers into its own package, modeled on the clean structure of pi-nvidia.

Setup

export CLOUDFLARE_API_KEY=...        # API token from dash.cloudflare.com
export CLOUDFLARE_ACCOUNT_ID=...     # Account ID from the dashboard URL

Install like any pi extension:

npm install pi-cloudflare-workers-ai

Then restart pi so the cloudflare-workers-ai provider is loaded. The account id is injected into the request URL at request time (it cannot live in a static base URL), so both env vars are required before any call — pi will show a clear error otherwise.

Two API surfaces

Cloudflare exposes two surfaces, both handled here:

• OpenAI-compatible chat — POST /accounts/{account_id}/ai/v1/chat/completions. The extension delegates these to pi-ai's built-in openai-completions core (same compat path pi uses for OpenAI-compatible servers), so streaming, tool calls, and thinking all work.
• "Run a model" microservices — POST /accounts/{account_id}/ai/run/{model_id} for image / TTS / ASR. These are not chat completions, so they have dedicated native streamers.

Auth is CLOUDFLARE_API_KEY as a Bearer token; CLOUDFLARE_ACCOUNT_ID goes in the URL path.

Free-tier models

The curated list is the verified free set (10,000 neurons/day, UTC reset). The five paid frontier models (deepseek-v4-flash/pro, glm-5.2, kimi-k2.6/k2.7-code) require billing and are intentionally not registered.

Model	Type	Reasoning	Context	Max Out
@cf/openai/gpt-oss-120b	chat	✓	128K	64K
@cf/openai/gpt-oss-20b	chat	✓	128K	64K
@cf/meta/llama-3.3-70b-instruct-fp8-fast	chat	✓	24K	8K
@cf/qwen/qwen3-30b-a3b-fp8	chat	✓	32K	8K
@cf/qwen/qwen2.5-coder-32b-instruct	chat	✓	32K	8K
@cf/google/gemma-4-26b-a4b-it	chat	✓	128K	64K
@cf/zai-org/glm-4.7-flash	chat	✓	128K	64K
@cf/deepseek-ai/deepseek-r1-distill-qwen-32b	chat	✓	80K	64K
@cf/black-forest-labs/flux-1-schnell	image	—	—	—
@cf/deepgram/aura-2-en	tts	—	—	—
@cf/deepgram/aura-2-es	tts	—	—	—
@cf/openai/whisper	asr	—	—	—
@cf/deepgram/nova-3	asr	—	—	—

Commands

/cloudflare-models [chat|image|tts|asr]

Per-model capability table; the optional filter narrows to one type.

/cloudflare-models
/cloudflare-models image

/cloudflare-usage

Opens the Cloudflare dashboard in your browser. There is deliberately no data table — Cloudflare exposes no usage/quota API; Workers AI usage (neurons/day) is visible only in the dashboard UI.

Notes

• Generated images are saved under .pi/generated-images/ and reported as a clickable link (OSC 8 hyperlink); TTS audio under .pi/generated-audio/; transcripts under .pi/generated-transcripts/.
• The model id contains / and ., which are unsafe in filenames, so they are sanitized when writing generated media.
• If you also use pi-cn-free-model-providers, it registers its own cloudflare provider (chat + native media). Load one or the other to avoid an id clash.