François Kiene

Indie dev. Building llmtrim, an open-source proxy that cuts your LLM bill without busting the prompt cache. Rust, local-first.

[François Kiene @kiene_f]

You're paying full price for every token Claude Code sends, the first time it sends it. The cache discount only kicks in on the second pass.

That gap cost me ~$460 before I built a fix: llmtrim, an open-source proxy that compresses those tokens ~68% without busting the cache.

[François Kiene @kiene_f]

@aoagents in

[François Kiene @kiene_f]

@azrulrhm Thanks for the shoutout. I built llmtrim. If you run into anything, ping me anytime.

[azrul gpt @azrulrhm]

Try this, llmtrim. Similar to rtk, it claims to cut down noise in your tool calls, reducing the tokens sent to your LLM provider.

Initial test looks ok, but I am not yet seeing much quota increase. No harm trying.

[François Kiene @kiene_f]

@semihlaki Congrats on TAP, and thanks for the shoutout. Similar thing happened to me with headroom: found out about it after finishing my own product, mine ended up more performant, smaller, and faster. Would love to have you contribute to llmtrim sometime.

[François Kiene @kiene_f]

Was waiting for this! Here's mine: llmtrim-herdr, trims wasted tokens from every agent request.
https://t.co/jx73vDdtII

[herdr @herdrdev]

herdr 0.7.0 is out, and it's a major one: it introduces plugins!

the idea is simple: herdr stays lean, and everything custom gets extended through plugins. shareable, scoped, built however you want, to fit your own flow.

with this release we're also shipping a few examples of what the plugin system can do. first up: a telegram plugin.

herdr already controls your agents and knows their status, so the plugin just hooks into agent events and pings telegram the moment one needs you.

notification lands → `herdr --remote` or ssh from your phone → straight back to the agent that needs you.

[GitHubDaily @GitHub_Daily]

用 Claude Code、Codex 辅助写代码，有时候对话会发送大段报错日志和重复 JSON 结果，导致 Token 消耗大幅上涨。

偶然在 GitHub 上发现 llmtrim 这个项目，正好解决了这个问题，挺实用的。

它在本地构建了一层代理，先把请求里的废话压掉再发给模型，保持了回答质量不变，账单大幅缩水。

内置十种智能压缩策略，能折叠成百上千行的重复日志，也能把臃肿的 JSON 数组转码成紧凑格式。

GitHub：https://t.co/kQ9FCSuLA6

全程在本地机器上运行，无需加载额外模型，单次调用延迟极低，数据隐私非常有保障。

安装后不用改任何工具配置，支持 Claude Code、Codex、Cursor 等主流 AI 编程工具，也能当命令行工具或代码库直接调用。

[François Kiene @kiene_f]

@GithubProjects Improving llmtrim: a local proxy that strips wasted tokens out of every LLM request. Same answers, smaller bill: −31% input, −74% output. One static binary, ~5 ms per call, no model to load.

https://t.co/GjLwCYHOaa

[François Kiene @kiene_f]

@Marco_Ramilli Maker here, thanks for the share! That 31% is just input tokens. Output drops ~74%, so the round-trip bill is about 59–66% cheaper depending on the model. Lossless and re-checked every request, so it can never make your bill worse.

[Marco Ramilli @Marco_Ramilli]

🤖 llmtrim
⭐ 76 stars

Slash your LLM bill by up to 31% with zero loss in quality. This local proxy trims wasted tokens from prompts and code before they hit the API.

🔗 https://t.co/iQVs2OdXRT

#AI #MachineLearning

[François Kiene @kiene_f]

@TugserOkur Süper, kurulum tek komut. Takılırsan yaz, yardımcı olurum.

[François Kiene @kiene_f]

@RoundtableSpace Same space, different bet: llmtrim is a pure algorithm, no model to download or run, so you can audit every line. Lossless and re-measured every request, so it can never inflate your bill or break a call.

https://t.co/GjLwCYHOaa

[François Kiene @kiene_f]

@tonysimons_ Same space, different bet: llmtrim's pure algorithm, no model to download or run, so you can audit every line of it. Trades some compression for that. https://t.co/GjLwCYHOaa

[François Kiene @kiene_f]

@jeffreynichols @tonysimons_ Privacy is why I built llmtrim local-first. It just forwards to the LLM API you already call. The only other network call is an update check you can switch off. No telemetry. Open source, so audit every call yourself. https://t.co/GjLwCYHOaa

[Geek @geekbb]

在 LLM API 请求发出去之前，把里面浪费的 token 挤掉，省钱但不影响回答质量。

https://t.co/0z8NRJzw4H

[François Kiene @kiene_f]

@metrpay nice! https://t.co/emPxPKCLHU
DM me if anything's weird

[François Kiene @kiene_f]

@metrpay Haven't tried it so can't judge. It's a different axis though: they discount the price per token, llmtrim cuts how many you send, so they'd even combine. Only thing I'd check is being ok with your traffic going through a third party, mine stays on your machine

[François Kiene @kiene_f]

@robj3d3 nice stack. effort routing saves the thinking side, but tool output still rebills full price every turn. I trim that with a proxy I built (llmtrim, open source), -69% on my Fable traffic. stacks with yours

[François Kiene @kiene_f]

@metrpay been there. most of that is billed full price, the cache discount never sees it. I run Fable through a little proxy I built (fkiene/llmtrim, open source), saved $293 on my own Fable traffic so far

[François Kiene @kiene_f]

Repo: https://t.co/emPxPKCLHU (AGPL-3.0)

npm install -g @llmtrim/cli && llmtrim setup

Tool output shrinks 93-98%, replies -74%. Every compression stage is re-counted with the provider's tokenizer and reverted if it doesn't save: worst case is a no-op.

[François Kiene @kiene_f]

You're paying full price for every token Claude Code sends, the first time it sends it. The cache discount only kicks in on the second pass.

That gap cost me ~$460 before I built a fix: llmtrim, an open-source proxy that compresses those tokens ~68% without busting the cache.