AI Runtime Layer

Infrastructure for agentsthat run themselves.

Two lines of code give your agents a runtime that understands their behavior - routing, compressing, and recovering automatically, in-process, in under 2ms.

<2msadds no noticeable delay to any request
12+providers auto-patched, out of the box
98%of requests get the right model automatically
live-request-trace
00ms
Request arrives
18,400 tokens
Model: GPT-4o
02ms
LoopGuard check
No recursion detected
03ms
Context compressed
18,400 → 2,100 tokens
04ms
Cost estimated
$0.32 → $0.07
05ms
Route selected
GPT-4o → GPT-4o-mini
06ms
Policy approved
Within $0.10 limit
07ms
Execute model
Request forwarded
Live RequestAgent: Research Assistant
Status
LoopGuard
Compression
Routing
Budget Check
Original Cost
$32.13
Final Cost
$7.02
Saved
78%
Three Capabilities

One layer.
It understands what your agents are doing.

Most infrastructure is passive - it logs, reports, waits for you to look at a dashboard. Aurex is active. It sits in the request path itself and makes a decision on every call before it goes out.

01

Behavioral Awareness

Understands the shape of what your agents are doing - matching tool calls, context growth, and semantic patterns across a run. It's how Aurex tells a genuinely new step from a repeated one.

F1: 0.72hybrid structural + semantic detection
Read docs →
02

Adaptive Routing

Every call is scored for complexity and routed to the model that fits it - a frontier model when the task needs it, a smaller one when it doesn't. Compression and caching happen the same way, automatically.

98%cost recovered without manual prompt review
Read docs →
03

Provider Continuity

Aurex knows the health of every provider in your stack in real time and reroutes around a bad one without you writing failover logic.

0msadded latency vs routing through an external gateway
Read docs →

How It Works

Drop Aurex between
your app and your LLMs.

Aurex is a thin SDK layer, not a gateway you route through. It intercepts calls in-process, decides in under 2ms, and lets your code stay exactly as it is. Works with LangChain, LlamaIndex, Vercel AI SDK, and every major provider out of the box.

Your Application
Web App
Workers
Your infrastructure
AI Frameworks
LangChain / AI SDK
Routes each request
Intelligence Layer
AUREX SDK
LoopGuardActive
CompressorActive
Cost AuditorActive
Runs in-process
LLM Providers
OpenAI
Anthropic
Gemini / Ollama
Final execution
Aurex intercepts calls to all major providers0.5–2ms overhead12+ providers auto-patchedSHA-256 only - zero prompt egress
OpenAI
Anthropic
Google Gemini
LiteLLM
LangChain
LlamaIndex
Vercel AI SDK
Ollama
HuggingFace
AWS Bedrock
CrewAI
AutoGen
OpenAI
Anthropic
Google Gemini
LiteLLM
LangChain
LlamaIndex
Vercel AI SDK
Ollama
HuggingFace
AWS Bedrock
CrewAI
AutoGen
OpenAI
Anthropic
Google Gemini
LiteLLM
LangChain
LlamaIndex
Vercel AI SDK
Ollama
HuggingFace
AWS Bedrock
CrewAI
AutoGen
Works with most tools

Start preventing waste today.

Two lines of code. All four pillars running in-process. No data leaves your machine by default. See your Token Credit Score in seconds.

Need deeper optimization? We specialize in prompt optimization, tool definition refinement, and custom model routing.

No credit card required • 100% local processing • Cloud sync is opt-in

Command Palette

Search for a command to run...