DEV Community

Dor Amir
Dor Amir

Posted on

Using NadirClaw with Cursor

#cursor #ai #llm #opensource

Using NadirClaw with Cursor

Cursor is a VS Code fork built for AI-assisted coding. It supports GPT-5, Claude Sonnet 4.5, Gemini, and custom models. Problem: if you're using it 8 hours a day, those API calls add up fast.

NadirClaw cuts those costs 40-70% by routing simple prompts to cheap models and complex ones to premium. Here's how to set it up.

I'm Dor, author of NadirClaw. This is the setup I actually use.

Prerequisites

  • Cursor installed (https://cursor.sh)
  • Python 3.10+
  • API keys for at least one LLM provider (Anthropic, OpenAI, or Gemini)

Install NadirClaw 

curl -fsSL https://raw.githubusercontent.com/doramirdor/NadirClaw/main/install.sh | sh
Enter fullscreen mode Exit fullscreen mode

Or via pip:

pip install nadirclaw
Enter fullscreen mode Exit fullscreen mode

Configure Your Models

Run the interactive setup:

nadirclaw setup
Enter fullscreen mode Exit fullscreen mode

This walks you through:

  • Which providers you want to use
  • Your API keys
  • Which model handles simple prompts
  • Which model handles complex prompts

Example config (what I use):

  • Simple: gemini-2.5-flash (free tier, fast)
  • Complex: claude-sonnet-4-5 (premium, best quality)

Save this to ~/.nadirclaw/.env:

GEMINI_API_KEY=AIza...
ANTHROPIC_API_KEY=sk-ant-...
NADIRCLAW_SIMPLE_MODEL=gemini-2.5-flash
NADIRCLAW_COMPLEX_MODEL=claude-sonnet-4-5
Enter fullscreen mode Exit fullscreen mode

Start NadirClaw 

nadirclaw serve --verbose
Enter fullscreen mode Exit fullscreen mode

NadirClaw runs on http://localhost:8856. Keep this terminal open.

Configure Cursor

Open Cursor settings (Cmd+, or Ctrl+,), then:

  1. Go to Models → Model Settings
  2. Select OpenAI API (or add a custom provider)
  3. Set:
    • Base URL: http://localhost:8856/v1
    • API Key: local (NadirClaw doesn't require auth for local use)
    • Model: auto

Cursor will now route all requests through NadirClaw.

What Happens

Every prompt you send gets classified in ~10ms:

  • Simple prompts (docstrings, quick fixes, "what does this do?") → routed to Gemini Flash
  • Complex prompts (refactoring, architecture, multi-file edits) → routed to Claude Sonnet

NadirClaw handles:

  • Streaming (works like normal Cursor chat)
  • Tool use (codebase search, terminal commands)
  • Multi-turn conversations (sticks to the same model)
  • Context window overflow (auto-swaps to a larger model if needed)

Real Numbers

Here's a week of my actual Cursor usage (tracked via nadirclaw report):

Before NadirClaw:

  • 847 requests, all to Claude Sonnet 4.5
  • Prompt tokens: 621,400
  • Completion tokens: 284,100
  • Cost: $127.30

After NadirClaw:

  • Simple tier (64%): 542 requests to Gemini Flash → $8.20
  • Complex tier (36%): 305 requests to Claude Sonnet → $54.80
  • Total: $63.00
  • Savings: $64.30 (50%)

Your mileage will vary based on what you build and which models you choose.

See Your Stats 

# Full report
nadirclaw report

# Last 24 hours
nadirclaw report --since 24h

# Live dashboard
nadirclaw dashboard

# How much you saved
nadirclaw savings --since 7d
Enter fullscreen mode Exit fullscreen mode

Cost Optimization Tips

Maximum Savings (65%+) 

NADIRCLAW_SIMPLE_MODEL=ollama/llama3.1:8b  # free, runs locally
NADIRCLAW_COMPLEX_MODEL=claude-sonnet-4-5
Enter fullscreen mode Exit fullscreen mode

Requires Ollama installed. Simple prompts cost nothing.

Balanced (50-60% savings) 

NADIRCLAW_SIMPLE_MODEL=gemini-2.5-flash  # free tier
NADIRCLAW_COMPLEX_MODEL=claude-sonnet-4-5
Enter fullscreen mode Exit fullscreen mode

This is what I use. Gemini's free tier covers most simple prompts.

All-Claude (30-40% savings) 

NADIRCLAW_SIMPLE_MODEL=claude-haiku-4-5
NADIRCLAW_COMPLEX_MODEL=claude-sonnet-4-5
Enter fullscreen mode Exit fullscreen mode

Keeps you in the Anthropic ecosystem but still saves on simple prompts.

Troubleshooting

Cursor says "API key invalid"

Make sure NadirClaw is running (nadirclaw serve). The API key field should be exactly local.

All prompts route to complex model

Check classification with:

nadirclaw classify "Add a docstring to this function"
Enter fullscreen mode Exit fullscreen mode

If the confidence is low, try lowering the threshold:

NADIRCLAW_CONFIDENCE_THRESHOLD=0.04 nadirclaw serve
Enter fullscreen mode Exit fullscreen mode

Model not found

NadirClaw uses LiteLLM under the hood. Check supported models at https://docs.litellm.ai/docs/providers.

For Gemini, NadirClaw calls the native API directly (faster, better streaming).

Source

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

Cursor docs: https://docs.cursor.com

License

MIT

Top comments (0)