DEV Community

Cover image for I made my coding agent check for discounts before it pays for anything
Nir Feinstein
Nir Feinstein

Posted on Fully Autonomous

I made my coding agent check for discounts before it pays for anything

#ai #claude #opensource #productivity

Coding agents now pick SaaS plans, APIs, hosting and domains for us. They follow instructions well and almost never ask whether there is a cheaper way to get the same result.

Claude Code with frugal installed, answering a request to upgrade 8 people to Slack Pro with a sourced table of cheaper options

The 144 dollar example

Ask Claude Code: "Our team of 8 needs Slack Pro. Walk me through upgrading." You get correct steps for monthly billing. Annual billing is 144 USD a year cheaper, nonprofits get Pro free, and schools get 85 percent off. None of that came up.

What frugal does

frugal is an Agent Skill plus a tiny optional hook. Whenever a task involves paying, the agent:

  1. Sizes the check to the amount: one line for a 20 USD purchase, a short table for a 500 USD plan, a negotiation plan for a contract.
  2. Climbs a ladder, cheapest first: don't pay (free tier, open source, something you already own), pay less for the same thing (plan, seats, annual billing, refurbished, timing), programs (startup, student, open-source, nonprofit), codes and stacking, ask (sales discounts, retention offers), switch (a cheaper alternative).
  3. Verifies on the live page and labels every number verified, reported or unverified. It never invents a promo code.
  4. Reports in a few lines, then carries on with the task. It never pays, enters card details or accepts terms by itself.

It also covers the money the agent spends while coding: picking a model for an LLM feature (batch APIs, prompt caching, smaller models), choosing hosting, or adding a paid SaaS dependency.

Did it help?

I ran eight everyday requests through claude -p from an empty folder, with and without the plugin, and had a separate model judge both answers blind on five criteria: saving found, sourcing, honesty, helpfulness and length. Answers with frugal averaged 9.8 out of 10, against 5.5 without. Without it, Claude Code declined most shopping questions as out of scope. With it, it found the 199 USD AirPods price against a 249 USD list price, the 2-year minimum term on .ai domains, and the real cost of a dropshipping stack after the 1 USD trials end.

It is a small sample with a model judge, so treat it as a sanity check rather than a benchmark.

Try it 

npx skills add fainir/frugal
Enter fullscreen mode Exit fullscreen mode

Or in Claude Code:

/plugin marketplace add fainir/frugal
/plugin install frugal@frugal
Enter fullscreen mode Exit fullscreen mode

It works in Claude Code, Codex, Cursor, Gemini CLI, Copilot and OpenCode. The code, the evals and the program catalog (checked against official pages this week) are on GitHub: https://github.com/fainir/frugal

I would love to hear where it gets the numbers wrong.

Top comments (0)