NadirClaw: saving on LLM requests with smart prompt routing

NadirClaw is a system for intelligent LLM request routing that classifies prompts as simple or complex directly on the user's device, without sending data to servers. It then automatically selects the appropriate model — inexpensive for simple tasks, powerful for complex ones. Result: significant savings on API costs without loss of quality.

How Routing Works

NadirClaw works in three stages. First, a local classifier analyzes the incoming prompt and determines whether it's a simple or complex request — without making external API calls. This is the key moment: classification happens on the client side, which avoids unnecessary expenses. The prompt is never sent anywhere; it remains private and is processed locally.

Then the system selects the appropriate model. Simple requests are routed to more budget-friendly options, such as Gemini 1.5 Flash, while complex analytical or creative tasks go to more powerful versions, like Gemini 2.0 Pro. Developers can customize routes and classification thresholds to their needs — set at what complexity level to switch to an expensive model. This is flexibility that fixed pricing models don't have.

The third stage is executing the request in the selected model and returning the result. The entire process takes milliseconds, and the user gets an answer almost instantly. Meanwhile, routing analytics stay on the user's side and aren't collected centrally, which adds privacy.

Where We Save

Economy is the main advantage of this approach. Most applications get mixed traffic: some requests require complex processing, but most are simple and routine. If you send everything to one powerful (and expensive) model, costs grow linearly. NadirClaw solves this problem:

• Simple requests (word definitions, JSON parsing, brief summaries) cost 10 times less
• Local classification — zero costs for identifying task type, without involving an LLM
• Large-scale applications — if 70–80% of tasks are simple, overall costs drop by a third or more
• Long context caching — works equally well with cheap and expensive models
• No redundant API calls — only necessary requests to paid services

You can use NadirClaw in two ways. First — embed it in your application through a Python library, which will route requests automatically in the background. Second — experiment via CLI to understand which classification thresholds work best for your scenario. Installation is minimal, setup takes minutes, and integration requires no changes to your application's core logic.

What This Means

In a world where LLM API costs grow with application scale, NadirClaw offers a practical way to optimize expenses. This is especially useful for systems with large volumes of simple requests — support chatbots, FAQ systems, text classification, content moderation, request processing.

Now developers have a tool to keep costs under control without sacrificing quality for complex analytical and creative tasks. This is a step toward more responsible and economical use of LLMs in production.