Question 1

Does Curtly work with every LLM, or only specific models?

Accepted Answer

Curtly works with any LLM. Optimization operates across two layers: dynamic input compression before network dispatch, and output efficiency directive injection (Curtly Brevity Engine). Whether you route to OpenAI, Anthropic Claude, Groq, OpenRouter, or custom upstream endpoints, Curtly optimizes both the prompt payload and the generated response.

Question 2

Will this alter or break my code snippets, SQL queries, or JSON schemas?

Accepted Answer

Never. Before input compression begins, an in-memory AST parser locks code fences, inline code, SQL statements, JSON schemas, URLs, and template variables into an immutable vault. Compression only purges conversational padding, repetitive scrape noise, and empty formatting. On the output side, Curtly directs models to output clean, exact code without sycophantic chatter or markdown boilerplate.

Question 3

Why does Curtly optimize both input prompts and model outputs?

Accepted Answer

Most token tools only compress input prompts. But generated output tokens cost 3x to 5x more than input tokens on modern LLMs. Curtly attacks both sides: (1) Input compression shrinks prompts and dynamic context by up to 60%. (2) Output efficiency injection stops models from generating repetitive pleasantries, apologies, and markdown fluff, cutting output tokens by 30-50%.

Question 4

How does Curtly interact with Anthropic and OpenAI Prompt Caching?

Accepted Answer

Prompt caching discounts static prefixes like fixed system instructions, which Curtly preserves untouched. Curtly targets dynamic context bloat that cannot be cached anyway: live web scrapes, dynamic database payloads, tool outputs, and conversational context.

Question 5

How much latency does Curtly add to my requests?

Accepted Answer

Under 2ms on standard runtimes. Because compression and directive injection are performed entirely in memory with no remote hops, overhead is negligible. By sending up to 60% fewer tokens over the wire and generating 30-50% fewer output tokens, your overall end-to-end response is significantly faster.

Question 6

Do you store, log, or train on prompt payloads?

Accepted Answer

Never. Payloads are processed strictly in memory and discarded the instant compression completes. No prompt content or generation outputs are ever persisted to disks, stored in databases, or used for model training.

Question 7

How do I integrate Curtly into an existing codebase?

Accepted Answer

Integration takes two minutes. Use Curtly as a drop-in OpenAI or Anthropic proxy endpoint with zero code changes, or use our lightweight TypeScript and Python SDKs. Simply point your client base URL to Curtly or wrap your messages payload.

Capability / Dimension	Curtly (Two-Stage)	LLMLingua	Manual Prompts
Optimization coverage	Input + output (two-stage)	Input only	Output only
Code & schema safety	Zero corruption guarantee	Frequent syntax breakage	Not applicable
Integration effort	1-line baseURL swap	Heavy Python + PyTorch dependency	Manual prompt copy-pasting
Hardware & compute	0 MB local compute required	Requires local GPU & 7B model	0 MB local compute required
Latency overhead	Under 15ms proxy routing	300ms to 800ms overhead	0ms overhead
Total cost reduction	55% to 70% total saved	30% to 40% saved	25% to 35% saved

Spend less on AI.
Ship more.

How Curtly reduces your production LLM bill

Safe Vault AST input pruner

Output Brevity Governor

2-line drop-in proxy

Why engineering teams and CTOs choose Curtly

Benchmark scenario inspector

Agent memory & state

Ready to cut your company's AI infrastructure costs?

Start free, scale as you grow

Free

Growth

Enterprise

Need custom high-volume quotas or tailored setups?

Straight answers. No fluff.

Ready to cut your token bills?

Spend less on AI.Ship more.