DEV Community

Dor Amir
Dor Amir

Posted on

Using NadirClaw with Claude Code

#ai #llm #claude #tutorial

Using NadirClaw with Claude Code

Claude Code burns through Sonnet 4.5 tokens faster than you'd expect. A typical refactoring session can rack up $2-3 in API costs. NadirClaw cuts that by 50-70% by routing simple tasks to cheaper models automatically.

How It Works

Claude Code sends every prompt to whatever model you configured. NadirClaw sits between Claude Code and Anthropic, classifies each prompt, and routes trivial stuff (doc lookups, simple edits, boilerplate) to Haiku while keeping complex reasoning on Sonnet.

You don't change how you work. The savings happen invisibly.

Setup

1. Install NadirClaw 

git clone https://github.com/doramirdor/NadirClaw.git
cd NadirClaw
npm install
Enter fullscreen mode Exit fullscreen mode

2. Configure Your API Keys

Create .env:

ANTHROPIC_API_KEY=sk-ant-xxx
PORT=3000
Enter fullscreen mode Exit fullscreen mode

3. Start the Proxy 

npm start
Enter fullscreen mode Exit fullscreen mode

NadirClaw runs on localhost:3000 by default.

4. Point Claude Code at NadirClaw

Instead of using Anthropic's API directly, configure Claude Code to use your local proxy.

Option A: Environment variable

export ANTHROPIC_BASE_URL=http://localhost:3000
claude-code
Enter fullscreen mode Exit fullscreen mode

Option B: Config file

Edit ~/.config/claude-code/config.json:

{
  "apiBaseUrl": "http://localhost:3000",
  "apiKey": "sk-ant-xxx"
}
Enter fullscreen mode Exit fullscreen mode

(Check Claude Code docs for exact config path on your OS.)

5. Use Claude Code Normally

That's it. Every request now goes through NadirClaw's classifier.

What Gets Routed Where

NadirClaw uses a 200-line logistic regression classifier trained on real coding prompts. Here's what typically happens:

Routed to Haiku (cheap):

  • Documentation lookups
  • Simple refactoring (rename variable, extract function)
  • Boilerplate generation (test stubs, type definitions)
  • Quick syntax fixes
  • File reads and scans

Routed to Sonnet (smart):

  • Architecture decisions
  • Complex debugging
  • Algorithm design
  • Refactoring with business logic changes
  • Anything ambiguous or multi-step

The classifier checks:

  • Prompt length and complexity
  • Question words vs imperative tone
  • Code structure mentions
  • Technical depth indicators

It's not perfect, but 10ms of overhead is negligible compared to model latency.

Monitoring

NadirClaw logs every request with routing decision:

[HAIKU]  "rename this variable to userId"
[SONNET] "refactor this auth flow to support OAuth2 and SAML"
[HAIKU]  "add JSDoc comments to this function"
Enter fullscreen mode Exit fullscreen mode

Watch the console while you work. You'll see most trivial tasks hit Haiku.

Expected Savings

Based on real usage:

  • Light sessions (mostly edits): 40-50% cost reduction
  • Heavy sessions (refactoring + new features): 50-70% reduction
  • Pair programming (lots of back-and-forth): 60%+ reduction

Your mileage varies based on how you use Claude Code. The more simple edits you do, the bigger the savings.

Gotchas

1. Local proxy required

NadirClaw runs locally. If you're on a flaky connection or frequently switching networks, the proxy might drop. Set a systemd/launchd service if you want it always-on.

2. Classification mistakes happen

Occasionally a complex task gets routed to Haiku and produces garbage. Just retry the prompt. The cost of one misroute is tiny compared to always using Sonnet.

3. Model version assumptions

NadirClaw defaults to claude-sonnet-4-5 and claude-haiku-4-5. If you use different models in Claude Code, override the routes in config.json.

Why This Beats Manual Rules

You could write routing rules yourself:

if (prompt.includes("refactor")) return "sonnet";
if (prompt.length < 50) return "haiku";
Enter fullscreen mode Exit fullscreen mode

But keyword matching breaks fast. "Refactor this entire auth module to support multiple providers" is not the same as "refactor variable names for clarity."

NadirClaw's classifier understands prompt semantics, not just keywords. It's trained on real coding task distributions, so it learns patterns you'd miss.

Source

Full disclosure: I built NadirClaw. This guide reflects my actual usage with Claude Code over the past month. Savings numbers are from my own API bills.

GitHub: https://github.com/doramirdor/NadirClaw

Issues/questions welcome.

Dor Amir

February 2026

Top comments (0)