Repo Tryouts · TRYOUT 01 · TRIM
llmtrim tryout: the full numbers
Tryout 01 ran llmtrim (MPL-2.0), a local proxy that compresses what your AI agent sends before it reaches the API.
What they claim
From the repo: about −31% input tokens and −74% output, measured live.
How I tested it
My own agent on a direct Anthropic API key, with the same three read-only, real-world tasks run with and without llmtrim. That’s 9 API calls in total.
| Measure | Without | With llmtrim |
|---|---|---|
| Input tokens | 251,703 | 210,108 (−16.5%) |
| Answers | same on all 3 tasks | same |
| Latency | — | +12 ms per call |
Verdict: LATER
The savings are real, but smaller than claimed on my workload. It’s also one more process between my agent and the API: if it stops, every call stops. I’d turn it on for one high-volume app in its safe mode before routing everything through it.
Try it yourself
Install it, run your own agent’s normal tasks once with it and once without, and compare the input-token totals from your provider’s usage page.