TEST 001 · llmtrim: −16.5% tokens, still off

What it is

llmtrim sits between your AI tool (Claude Code, Codex, Cursor, your own app) and the provider’s API. Before each request goes out, it compresses the bulky parts: tool output, chat history, code and JSON. Then it forwards the smaller request, and the reply comes back unchanged.

How I tested it

I routed my own AI agent through it on a direct Anthropic API key and ran the same three read-only, real-world tasks with and without it: 9 API calls in all.

Measure	Without	With llmtrim
Input tokens	251,703	210,108 (−16.5%)
Answers	Same on all 3 tasks
Extra latency	About 12 ms per call

Token counts through a proxy are estimates, and savings swing a lot by session, so measure on your own fixed set of tasks. The author reports −31% input tokens across 112 A/B cases; those are the author’s numbers, on a different mix of requests.

Why it’s LATER for me

Every call runs through it, API key included. It doesn’t store keys, but it installs a private certificate so it can read requests in transit (scoped to AI API hosts only).
It’s a single point of failure. If the proxy breaks, every call breaks.
Most of my tokens were already cached. Cached input is billed at a fraction of the price, so my real money saved was small.

Worth trying if you pay full price per token and don’t use prompt caching much, for example a custom app or agent calling the API directly.

Safe setup, step by step

Start in safe mode: lossless compression only. The default auto mode saves more, but some of its stages are lossy. In the author’s own benchmark, one maths test dropped from 100% to 92%.

Measure first. Pick 3 real tasks you run often and note the tokens each one uses today.
Install it.
```
npm install -g @llmtrim/cli@latest
```
Set it up. llmtrim setup adds a private certificate in ~/.llmtrim/, your shell’s HTTPS_PROXY setting and a login service. To keep it to one app, set the proxy only in that app’s environment.
```
llmtrim setup
```
Switch to safe mode. Or add preset = "safe" to ~/.config/llmtrim/config.toml.
```
export LLMTRIM_PRESET=safe
```
Check it’s capturing traffic.
```
llmtrim doctor
```
Re-run your 3 tasks and compare tokens and answers. llmtrim status shows the savings.
Keep it only if the saving is clear. Otherwise remove everything in one command:
```
llmtrim uninstall
```

Commands checked against the project’s README on 2026-11-02. GitHub Copilot doesn’t work through it (certificate pinning).

llmtrim: −16.5% tokens, still off