Skip to main content
Curtly is a high-performance AI Code & Quota Governor that cuts total LLM API costs by 50% to 70% through two-stage in-memory optimization:
  1. Input Prompt Compression (Safe Vault™): Deterministic AST parsing prunes conversational filler, verbose scaffolding, and scrape padding by up to 60% with sub-2ms latency and 0.00% syntax drift.
  2. Output Efficiency Injection (Curtly Brevity Engine): Injects concise generation directives into model context, cutting expensive output tokens by 30% to 50% (and up to 75%+ on code generation) with zero loss in output quality or reasoning capability.

​
Why Dual-Engine Optimization?

Most token optimization tools only focus on compressing input prompts. However, modern frontier LLMs (Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, Grok 4.7, Gemini 3.8 Flash, DeepSeek-V4.1-Flash) price output tokens 3x to 5x higher than input tokens, with Claude Fable 5.1 demonstrating the highest net cost savings under Curtly’s dual-engine optimization. By governing both ends of the inference lifecycle, Curtly delivers the highest net cost reduction in the industry while accelerating end-to-end response time.

​
Core Pillars

Safe Vault™ AST Engine

Guarantees 0.00% syntax drift. Code fences, JSON schemas, SQL queries, URLs, and template variables are locked in an immutable vault before compression begins.

Output Brevity Governor

Instructs models to omit sycophantic pleasantries, apologies, and markdown padding, delivering dense, exact answers.

Sub-2ms In-Memory Speed

Executes in volatile memory with zero local PyTorch or GPU requirements. Adds negligible latency to inference pipelines.

​
Architectural Comparison

​
Deployment & Integration Models

Curtly supports three frictionless deployment modes:
  1. Drop-in Proxy (Recommended): Swap your OpenAI or Anthropic baseURL to https://curtly.dev/v1 with zero application code changes.
  2. Official SDK (curtly): Use our TypeScript / Node.js SDK via npm install curtly or Python client with drop-in wrapper support.
  3. REST API & MCP: Call POST /api/v1/compress directly or connect our MCP server to Cursor, Claude Code, and Windsurf.

​
Security & Privacy Guarantee

  • Zero Data Retention (ZDR): Prompts and completion streams are processed strictly in volatile RAM and immediately discarded.
  • No Model Training: Your data is never stored on disk, written to persistent databases, or used for model training.
  • TLS 1.3 & Scoped Keys: All transit is encrypted end-to-end with granular API key scopes.