70% of API spend is duplicate inference <1ms cache hit response time 7 days free trial

70% of every AI API dollar
pays for identical work.

Your system prompt, few-shot examples, and context window are the same bytes, same request, same response — billed at full rate every time. CachePilot sits inline and returns cached responses for duplicates.

⚡ No signup required — send your provider key, get a CachePilot key back
curl — auto-provision
# First request: no CachePilot key needed — just your provider key curl https://cachepilot.koaw.workers.dev/v1/chat/completions \
  -H "x-provider-key: sk-your-openai-key" \
  -d '{ "model": "gpt-4", "messages": [{"role":"user","content":"hello"}] }'

# Response includes X-CachePilot-Key header — use it for subsequent requests < x-cachepilot-key: sk-cache-22b8ea87-6ae2c11c

Or create an account in 10 seconds:

No credit card. First $20 savings free. Pay 20% of what you save after that.

openai client
// Before: duplicates cost full price every time const openai = new OpenAI({ apiKey: "sk-..." });

// After: duplicates return from cache → $0 provider cost const openai = new OpenAI({ baseURL: "https://cachepilot.koaw.workers.dev/v1", apiKey: "sk-cache-xxx-yyy" });
How much could you save? →
0%
typical duplicate rate
<1ms
cache hit latency
0
cache layers
0%
our fee on savings

One line. No SDK. No config.

Change the base URL in your AI client. That's it. Every duplicate request stops costing you.

❶

Route through CachePilot

Change one string — baseURL points to CachePilot instead of the provider. No signup needed — send your provider key as x-provider-key and your CachePilot key is returned in the response header. Your SDK, auth, streaming, and response format stay identical.

❷

Duplicates never reach the provider

Every request is hashed and checked against our cache stack (L0 memory → L1 edge cache → L1.5 normalizer → L3 prompt optimization). On hit: response returns in <1ms, $0 provider cost.

❸

You only pay for unique requests

We track what you would have paid the provider vs. what you actually paid. Our fee is 20% of the difference. If you save nothing, you pay nothing — not a cent.

OpenAI • GPT-4o, GPT-4, o1 Anthropic • Claude Sonnet, Haiku OpenRouter • 300+ models Bring your own API key

Why 70% is the floor, not the ceiling

Real-world AI apps repeat the same requests far more often than developers realize.

📊

System prompts dominate

Most production apps attach the same system prompt (instructions, tone guardrails, output schema) to every request. That's 500–2000 tokens of identical text — billed as fresh input each time.

🔄

Background jobs repeat

Batch processing, data enrichment, nightly summaries — these hit the exact same prompt thousands of times. With no cache, every run is a full-price API call. With CachePilot, 100% of duplicates cost $0.

👥

Multi-tenant sharing

If ten users ask the same question, caching returns the same answer immediately — and only the first request ever hits the provider. For SaaS apps, this alone can cut costs by 60–90%.

📏

Prompt caching gaps

Provider prompt caching helps on multi-turn conversations, but doesn't cover exact duplicate requests across users, sessions, or time. That's the gap CachePilot fills — and where most of the waste lives.

Your actual savings are one signup away

The cache stack

Five layers, from nanoseconds to seconds, designed to maximize hit rate.

⚡

L0 — In-memory

Hot responses held in process memory with LRU eviction and 30-second TTL. Nanosecond lookup for the most frequent duplicates.

🌐

L1 — Edge cache

Cloudflare Cache API stores full responses at the edge. Shared across regions, tenant-isolated by hashed API key. Millisecond lookup, one-hour TTL.

🔍

L1.5 — Normalizer

Deterministic cache key generation that's blind to whitespace, key ordering, and formatting differences. "Hello" and " hello " produce the same cache key.

🧠

L3 — Provider optimization

Anthropic prompt caching markers injected automatically on session boundaries. Reduces re-processing of conversation history by up to 90%.

🔐

Tenant isolation

Every cache key is prefixed with a SHA-256 hash of the customer's API key. Your users' data never leaks across tenants. Passthrough mode available for sensitive requests.

📋

Session-aware

Multi-turn conversations are detected automatically. Message history is hashed and compared — no raw content stored. Continuations get provider-level prompt caching for the stable prefix.

Pricing that matches your savings

No subscription. No minimum. You pay only when the cache saves you money.

First 7 days free — no charge during trial
20%
of calculated savings

Save $100 → you pay $20. Save $1,000 → you pay $200.
Save $0 → you pay $0. Forever.

  • Only pay on what you save
  • First week free for all new signups
  • No subscription or minimum spend
  • Cancel anytime, keep your API key
  • Transparent per-request cost breakdown
Your monthly AI spend: $_______
With 65% cache hit rate → you save $_______
Your CachePilot fee: 20% of savings

FAQ

Questions developers ask before changing their base URL.

Do I need to sign up first?▼
No. Send your first request with your provider API key in the x-provider-key header — no CachePilot key needed. We auto-create an account and return your sk-cache-xxx key in the X-CachePilot-Key response header. Use that key (without x-provider-key) for subsequent requests. You can also sign up with your email above for a pre-generated key.
What happens during the 7-day trial?▼
Your fee is waived on all savings for the first 7 days after signup. We still cache your requests and calculate what you would have paid — but the bill is $0. After the trial, you're charged 20% of savings going forward. If you saved $0 during the trial, you owe nothing after it either.
Do I have to change any code?▼
One line: replace baseURL in your OpenAI/Anthropic client with https://cachepilot.koaw.workers.dev/v1. Your SDK, authentication, streaming, and response format all stay exactly the same. Change it back anytime — there's no lock-in.
What if my app has no duplicate requests?▼
Then you pay $0 forever. We calculate savings per-request: what the provider would have charged minus what you paid through us. If every request is unique, your savings are $0 and our fee is $0. Most apps have far more repetition than they expect — system prompts, few-shot examples, background jobs, shared user queries.
How is savings calculated?▼
We track each request at provider pricing (model, input tokens, output tokens). A cache hit means the provider was never called — $0 cost. A cache miss means the request was forwarded to the provider and their standard rate applies. Savings = baseline − actual. Our fee = 20% of cumulative savings. Everything is visible in the dashboard.
Which providers and models are supported?▼
Any OpenAI-compatible API: OpenAI (GPT-4o, GPT-4, o1, o3), Anthropic (Claude Sonnet 4, Haiku), and all 300+ models on OpenRouter. You bring your own provider API keys and pass them via the x-provider-key header. We never handle provider authentication.
Are my prompts stored or logged?▼
No. We store only SHA-256 hashes of message sequences for session detection — never the raw content. Cache keys are deterministic hashes of the normalized request. For sensitive requests, send x-cachepilot-mode: passthrough and the request bypasses all caching entirely.
Does streaming work?▼
Yes. Streaming responses pass through CachePilot transparently. They are not cached as full responses (since each stream is unique), but CachePilot injects provider-level prompt-caching markers on the stable prefix before forwarding — so you still get prefix caching savings from the provider (e.g., Anthropic's cached prompt tokens at 10% of input cost). Streaming gets provider-level prefix caching savings but not full-response caching.
What's your uptime guarantee?▼
CachePilot runs on Cloudflare Workers — the same edge network that powers millions of production sites. There's no tunnel, no SSH, no local machine involved. If Cloudflare is up, CachePilot is up. No planned downtime.