Summary

Three tools if you want the engine, a proxy if you want it automatic.

The MCP surface is the small door into a larger product: three tools that compress on request, with none of the CA-and-proxy setup the automatic path asks for. Two things hold either way — the default preset is quality-gated rather than lossless, so reach for `safe` when you need a byte-faithful round trip, and Anthropic and Gemini token counts are a BPE approximation flagged in `status`, because neither ships an exact public tokenizer.

What it is

A local proxy that trims LLM API traffic and forwards a smaller request, plus the same engine as a CLI, a library in six languages, and an MCP server. Compression runs on your machine and in-process, with no extra model call and nothing sent to the project. As an MCP server it is three tools — `llmtrim_compress`, `llmtrim_compress_text` and `llmtrim_stats`, the last reading the same ledger `llmtrim status` shows. Written in Rust (1.88+), MPL-2.0.

What you get

`llmtrim_compress` and `llmtrim_compress_text` — run a request or a block of text through the same engine the proxy uses, with no proxy involved
`llmtrim_stats` — the savings ledger, the same one `llmtrim status` reads
Client setup in one command: `llmtrim mcp install` for Claude Code, `llmtrim mcp install --print` to paste into anything else; the entry is the `llmtrim` command with the argument `mcp`
A proxy path for anything that honours `HTTPS_PROXY` — Claude Code, Codex CLI, Gemini CLI, Cursor, Cline, Roo, Aider and others. GitHub Copilot is out, because of certificate pinning
Presets chosen by traffic shape: `auto` (default, decides per request), `safe` (lossless input only), `aggressive` (max squeeze, quality-gated), plus `agent`, `code`, `rag`, `cache`, `reasoning` and `frugal` — set through `LLMTRIM_PRESET` or `preset` in `$XDG_CONFIG_HOME/llmtrim/config.toml`
The same engine as a library — `cargo add llmtrim-core`, `pip install llmtrim`, `gem install llmtrim`, `@llmtrim/js` over WASM, plus Kotlin and Swift packages
A savings view in the terminal — `llmtrim status`, with `--daily`, `--weekly`, `--monthly`, `--json` and `--csv` — and a menu-bar tray showing the same numbers
Published A/B numbers rather than claims: across 112 live cases, input tokens 71,031 → 49,062, output 25,843 → 6,628, round-trip cost $0.0365 → $0.0126, answer quality 78.9% → 82.2%

Requirements

For the MCP tools, the `llmtrim` binary on your PATH and nothing else — `npm install -g @llmtrim/cli@latest`, or Homebrew, Scoop, Cargo, the install script or the Docker image; the registry entry runs the npm package `@llmtrim/cli` as `llmtrim mcp` over stdio. Compression happens in-process, so there is no proxy to stand up and no API key stored. The proxy is a separate decision with a bigger footprint: `llmtrim setup` changes three things and `llmtrim uninstall` reverses all three — a name-constrained private CA in `~/.llmtrim/`, `HTTPS_PROXY` plus CA trust in your shell environment, and a login service. Prompts never touch disk; recoverable tool results stay in bounded daemon RAM for five hours by default and disappear on restart.

Setup effort

One command — llmtrim mcp install