The seatbelt for vibe-coded AI apps
Your AI app doesn't have a seatbelt. Prune is one line away.
One leaked key or a runaway loop is all it takes. Prune vaults your key, caps your spend, signs a receipt for every call, and automatically cuts your costs.
I personally walk every beta user through connecting their first key.
Private beta access. We onboard in small batches. Prompt bodies are processed statelessly and are not used for training. Provider keys are encrypted at rest (AES-256).
No spam, ever. Already invited? Log in
from prune import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-5.5",
messages=[{"role": "user", "content": prompt}],
)Toggle what Prune applies — watch the receipt update
Signed proof on every response
About 33% cheaper on this call — and cryptographically signed, so the numbers can be verified, not just trusted.
Engineering debugs with it. Finance reports with it. Auditors can verify it.
Full receipt (for your logs)
{ "model": "claude-opus-4-8", "choices": [{ "message": { "content": "{ ... }" } }], "usage": { "prompt_tokens": 684, "completion_tokens": 412 }, "prune_metadata": { "cache_hit": false, "cost_without_prune_usd": 0.460, "cost_with_prune_usd": 0.310, "unique_path_cost_saved_usd": 0.150, "cache_path_cost_saved_usd": 0, "tokens_saved": 3400, "optimizations_applied": [ "model_routing", "template_opt", "output_cap:512" ], "pii_shield": { "active": true, "fields_masked": ["email"] }, "receipt_signature": "kQ7vN2LxTz8mYw4pRc…hZ3w==", "receipt_signing_key_id": "prune-ed25519-2026-01" }}Your app
One base URL change
Prune
Vault · Cap · Receipt
Providers
OpenAI · Claude · GeminiOpenAI · Claude · Gemini · Bedrock
← a signed receipt returns with every response
Security by architecture
Built for trust, not badge-first marketing
Prune is in private beta and not yet SOC2-certified. We focus on clear system behavior instead of compliance theater.
- Provider keys are encrypted at rest with AES-256.
- Each customer's data is isolated by account and key scope.
- Prompt bodies are handled statelessly in transit and not stored for training.
- Policy checks run before requests are sent to external providers.
If you are evaluating Prune, evaluate what is stored, what is blocked, and what leaves your stack.
What actually runs through Prune
Works with
OpenAIAnthropic
Google Gemini
AWSOllama
OpenRouter
- +6 more
What you get
Vault. Cap. Receipt. Savings.
Your key never sits in your code
Provider API keys are locked in an AES-256-GCM encrypted vault, never exposed in your app or repo. If your codebase leaks, the key doesn't — and Prune also blocks personal data, secrets, and injection attempts before they reach OpenAI, Claude, or Gemini, and checks every response for signs of a leak.
Spend stops before it happens
Set a hard limit per key. Prune blocks the call before it's exceeded, not after — a runaway agent loop can't turn into a runaway bill.
A signed receipt on every response
Every call returns a breakdown of what it cost, which model was used, and which feature triggered it — cryptographically signed (Ed25519), so finance and auditors can verify the numbers instead of taking a dashboard's word for it.
Lower bills on top
You do not need repeat traffic to see savings. Prune routes to cheaper models when safe, trims verbose prompts, catches near-duplicate questions even when the wording changes, and caches exact repeats — savings pay for the product.
FAQ
Not another cache gateway
What is Prune?+
Prune is a managed AI gateway. Change one base URL so every call to OpenAI, Claude, Gemini, Bedrock, or OpenRouter goes through Prune — for unique-path cost control, Shield security, and a receipt on every response. Cache helps on repeats; it is not the product thesis.
How is Prune different from cache-focused AI gateways?+
Most AI gateways lead with cache hit rate. Prune leads with unique-path savings (routing, output caps, prompt cleanup) that apply when the prompt never repeats, plus Shield (vaulted keys, PII/secret checks, spend caps) and a machine-readable receipt. Cache is measured separately as upside.
How does Prune reduce LLM costs?+
On every call Prune can route to a cheaper model when safe, enforce output limits, and trim wasted prompt text. When traffic repeats, exact and semantic cache add extra savings. Receipts break out unique-path vs cache savings so you are not sold a vanity hit rate.
Does Prune only help when prompts repeat?+
No. The non-negotiable thesis is that Prune must still win on unique traffic. Caching and semantic reuse are multipliers on top, shown separately in prune_metadata.
Is Prune secure?+
Provider API keys are stored in an encrypted vault (AES-256-GCM). Shield can block PII, secrets, and runaway spend before the provider call. Prune is in private beta and is not SOC2-certified; evaluate architecture, not badge theater.
Where are provider API keys stored?+
Provider keys never live in your app or repo. They are stored in Prune’s AES-256-GCM encrypted vault, scoped per account and key. Your client only holds a Prune API key that authorizes calls through the proxy.
How does the proxy work?+
You point your existing OpenAI/Claude/Gemini SDK at Prune’s base URL (or use the Prune SDK). Prune vaults the provider key, applies spend caps and Shield checks, optionally optimizes the call, then forwards to the provider and returns the response with a signed receipt.
What is in a signed receipt?+
Each response can include prune_metadata: cost with and without Prune, unique-path vs cache savings, tokens saved, optimizations applied, and an Ed25519 receipt_signature with a signing key id — so finance and auditors can verify the numbers.
Which providers does Prune support?+
OpenAI, Anthropic (Claude), Google Gemini, AWS Bedrock, OpenRouter, and Ollama — via drop-in base URL or Python/TypeScript SDKs.
How do I verify a Prune receipt?+
Every response can include prune_metadata. When receipt signing is configured, paste the object at https://www.withprune.com/verify or use the SDK verify helpers against the published Ed25519 public key.
How long does it take to integrate?+
Typically under five minutes: create a Prune API key, vault your provider key, point your SDK at Prune, send one request, and read the receipt.
More detail: vs cache-first gateways · how it works