Show HN: Get more out of your Claude Code plan with VibeBooster

3 points

9 months ago

Seeing the recent adjustment of Claude Code limits made me think about how the incentives of Anthropic and users are not necessarily aligned for Claude code. Anthropic probably errs on the side of sending too many tokens, while some users might be okay with risking lower quality output for the sake of getting more out of their Claude code plan. So, I developed VibeBooster, which is a proxy that intercepts and shorten the requests that Claude Code makes to the Anthropic API.

By using a cheaper (or even free) LLM, we can summarize, compress, or even outright delete some extraneous tokens that likely wouldn't help Claude Code's output. There are some pretty egregious wastes of your precious tokens that I found by inspecting the requests that were made, and VibeBooster can avoid that waste! From my casual testing, I've been able to save about *40%* of the input tokens sent to the Anthropic API via Claude Code on large messages. Check out the README.md if you want to see some examples of token savings.

The project is open source, and bring your own key. I think there's potential to further prompt tune this, and also flip the philosophy of the product by trying to improve the performance of Claude Code rather than trying to save tokens. I don't feel this is a pure utility; it's also a research project into how Claude Code works, and an art piece demonstrating the weird state LLM-powered products are at this point in time.

Please let me know any feedback or suggestions you might have, thanks!