DEV Community

Maor Bril
Maor Bril

Posted on

The $25 Challenge: what can an agent build before the meter stops?

#opensource #llm #ai #cli

Claude Code is the best agent harness I've used. It also only talks to one provider, and it never answers the two questions you eventually ask: how much did that session cost, and can I run the cheap parts on a cheap model?

I built gremlord to fix both. It's a thin local router around the unmodified claude binary - same TUI, same tools - but the model behind it can be Anthropic, OpenAI, xAI, Ollama, or any OpenAI-compatible endpoint. Every token is priced as it streams, and a hard budget stops the next request at the cap instead of warning you after. Free, MIT licensed: https://github.com/gremlord/gremlord

Now I want to see what other people build with it, so I'm running a contest:

The rules

Build anything real with Claude Code running through gremlord, on any model you like. One constraint: total model spend for the build must be $25.00 or less, proven by your gremlord cost log.

  1. Set your cap before you start: gremlord budget set --daily 25. (That's a daily cap, so the contest receipt is what matters: gremlord cost --receipt --since YYYY-MM-DD from the day you started. That number plus any local hardware cost must be <= $25.00.)
  2. The build must do something - a tool, a fix, a feature, a demo. Presenting itself doesn't count.
  3. Using a passthrough or subscription-billed CLI delegation (Codex/Grok under your own login)? Disclose it. Unpriced subscription work is recognition-only unless you assign it a fixed rate.
  4. Running local models (Ollama, vLLM)? Price them by fixed hardware tier x active runtime - a 4-12B Q4 on a laptop is $0.10/hr, a 24GB GPU box is $0.35/hr, full table on the challenge page.
  5. Honor system on receipts. A faked log to win a README mention is a sad life.
  6. One entry per person. Entries close September 17, 2026 23:59:59 UTC.

Winning

Two winners, no panel. Community pick: most thumbs-up on your entry at cutoff. Founder's pick: most impressive build under budget, chosen by me, wins a $25 bounty. Both get a permanent slot in the challenge leaderboard and the gremlord README.

How to enter

  1. Install gremlord: https://github.com/gremlord/gremlord
  2. gremlord budget set --daily 25
  3. Build. Then gremlord cost --receipt --since YYYY-MM-DD.
  4. Reply to the challenge thread with your build, the receipt pasted in full, any disclosure, and your hardware tier + active runtime for local models: https://github.com/gremlord/gremlord/discussions/29

Full rules: https://gremlord.com/challenge/

The interesting part isn't the bounty - it's the constraint. $25 is about 5M tokens of a frontier model, or roughly 250 hours of a small local one. Those are very different builds. I honestly don't know which wins.

Window is open. Show me what you've got.

Top comments (0)