claude-skills/cloudflare/references/ai-gateway/features.md

Features & Capabilities

Caching

Dashboard: Settings → Cache Responses → Enable

// Custom TTL (1 hour)
headers: { 'cf-aig-cache-ttl': '3600' }

// Skip cache
headers: { 'cf-aig-skip-cache': 'true' }

// Custom cache key
headers: { 'cf-aig-cache-key': 'greeting-en' }

Limits: TTL 60s - 30 days. Does NOT work with streaming.

Rate Limiting

Dashboard: Settings → Rate-limiting → Enable

• Fixed window: Resets at intervals
• Sliding window: Rolling window (more accurate)
• Returns 429 when exceeded

Guardrails

Dashboard: Settings → Guardrails → Enable

Filter prompts/responses for inappropriate content. Actions: Flag (log) or Block (reject).

Data Loss Prevention (DLP)

Dashboard: Settings → DLP → Enable

Detect PII (emails, SSNs, credit cards). Actions: Flag, Block, or Redact.

Billing Modes

Mode	Description	Setup
Unified Billing	Pay through Cloudflare, no provider keys	Use cf-aig-authorization header only
BYOK	Store provider keys in dashboard	Add keys in Provider Keys section
Pass-through	Send provider key with each request	Include provider's auth header

Zero Data Retention

Dashboard: Settings → Privacy → Zero Data Retention

No prompts/responses stored. Request counts and costs still tracked.

Logging

Dashboard: Settings → Logs → Enable (up to 10M logs)

Each entry: prompt, response, provider, model, tokens, cost, duration, cache status, metadata.

// Skip logging for request
headers: { 'cf-aig-collect-log': 'false' }

Export: Use Logpush to S3, GCS, Datadog, Splunk, etc.

Custom Cost Tracking

For models not in Cloudflare's pricing database:

Dashboard: Gateway → Settings → Custom Costs

Or via API: set model, input_cost, output_cost.

Supported Providers (22+)

Provider	Unified API	Notes
OpenAI	openai/gpt-4o	Full support
Anthropic	anthropic/claude-sonnet-4-5	Full support
Google AI	google-ai-studio/gemini-2.0-flash	Full support
Workers AI	workersai/@cf/meta/llama-3	Native
Azure OpenAI	azure-openai/*	Deployment names
AWS Bedrock	Provider endpoint only	/bedrock/*
Groq	groq/*	Fast inference
Mistral, Cohere, Perplexity, xAI, DeepSeek, Cerebras	Full support	-

Best Practices

1. Enable caching for deterministic prompts
2. Set rate limits to prevent abuse
3. Use guardrails for user-facing AI
4. Enable DLP for sensitive data
5. Use unified billing or BYOK for simpler key management
6. Enable logging for debugging
7. Use zero data retention when privacy required