I would like to have deeper comparison with alternatives like rtk, which are already fast and written in rust, also the previous comments mentioned something that has been a know problem with rtk that it sometimes strips the thing that the llm needs (or expects, causing more work to need to happan not less)
None of these tools measure how effective they are...
It's a massive red flag to me when you could get decent data to see if your thing actually works, and they don't even attempt to...
Have the LLM use your tool, run it on several of the coding benchmarks. If you're stingy, run it on the ones that don't cost much.
Otherwise, I'm going to assume it doesn't actually work. If it did - Claude, Antigravity, Codex, Pi, or some major player would bundle tools like this into the CLI / harness.
AFAIK, none of the major players do. That's a sign to me these don't work in general.
I've tried building some tools specific to bug fixing. Intelligently feeding context massively helps smaller models. But, what I've found - surprisingly - is that a smaller, much better focused, including a lot of helpful data as well, has almost no impact on larger models compared to what they do by default.
You do save some tokens, though, which is what they're claiming - but not ~99%...
> otherwise, popular solutions would integrate the idea
None of the major players are incentivized to care about this, especially not over other opportunities. Why would you expect them to integrate it?
One of the biggest wins you can institute for your own codebase if you use agents is writing your own harness, by a huge margin. The defaults are fine, but you can do better.
They're incentivised because they're offering plans at a loss and/or pricing out potential customers. All these LLM companies are competing on accuracy and price.
If you're looking for an efficiency-focused harness, I had a pretty good time using the Dirac agent. The line-based anchors were slightly buggy though (this was a couple months ago) and would sometimes add the same line of code multiple times or leave an anchor in the output.
I don't think frontier model providers are going to be incentivized to invest in this much, yet. Once inference gets more competitive, sure. I haven't looked lately, but won't be surprised if tools like OpenCode do do what you're suggesting, though. Third-party coding harnesses ARE aligned to deliver this type of feature and optimization.
It's too hard to define what "works" even means in this case. Look at the example savings output. A lot of it is kubectl output.
Your suggestion to using coding benchmarks doesn't really capture the whole picture. I haven't seen a benchmark using kubectl.
> AFAIK, none of the major players do. That's a sign to me these don't work in general.
It's a lose/lose for major players. If it works well, it will lower their revenue. Also there's a high risk it'll significantly worsen results for some people, even if it improves results for others.
So often we will burn 20% of limit in a single ill conceived agent tool call that we're simply not going to be able to or want to be able to intercept. Where I see a tool like this being a real step forward is to add a decision point. it does not have to bubble up to hard-require user to provide permission, but it can let the LLM have an intermediate checkpoint to say that it's about to get blasted with 30k tokens and here is roughly the shape of it and do you wanna adjust or whittle it down if you know what you're looking for etc.?
There is definitely tons of value to extract from this line of thinking.
> I'm going to assume it doesn't actually work. If it did - Claude, Antigravity, Codex, Pi, or some major player would bundle tools like this into the CLI / harness.
The problem even attempting to develop a tool for the frontier model space is that the cost to run a statistically significant benchmark is almost certainly going to be over $100 - for a single model.
Unless something is like 25%+ more cost effective on Gemini for a task, I would not assume those savings are going to transfer to GPT.
If you need to run a test this expensive and slow for every release, hobbiests aren't going to do it.
And if you wanted any broadly specific improvements to coding like they all claim, the costs would be in the thousands per release even for a single for a single model.
And they almost certainly would not be eye popping.
If the models could be SUBSTANTIALLY better, Google and Anthropic and OpenAI wouldn't be finding that out from a hobbiest making wildly unscientific claims.
Yup, this is hitting it on the nose. But, despite the cost - the benchmark is the vital ingredient that cant be skipped. Otherwise, you don't know if what you're building is actually helping the agent rather than hindering it.
On the previous large benchmark run, i proved 40-50% cost reduction per correct answer.
I'm not sure why the vendors aren't using token filtering/compression more in their tooling, but perhaps they don't mind users feeding them more data and using more data.
You can't measure effectiveness, because you never know what kind of model will process your prompt. One request you might get full e.g. Opus and another they'll downgrade it to Sonnet or something more basic. I have this with "Opus 4.8" all the time.
In term of token saving performance, it should be on par with rtk since it is basically the same idea.
The major different is rtk bundled hundreds of filter logic and no room for user to adjust without maintaing user owned fork or opening the pull request while lowfat is using opposite architectural approach by removing almost all filter logic in the binary and seperate user filters as a plugin system
I've tried rtx and lean-ctx and these tools seem to end up confusing the agent more than helping. Any saving is irrelevant if the agent decides to work around the tool and makes even more calls than it would otherwise.
I don't know about cost saving, but if it's keeping the context size down I've had a lot better results using subagents to keep a higher order conversation clean for longer.
I looked into lean-ctx and decided not to use it. It has a very specific use case, and it's good when your interaction with the repository is read-only. When you want to edit, then the model has to read the whole file anyway. It's a cool tool, but it has a very narrow use case where it delivers the performance it claims.
The docs are missing any examples of what this does, instead showing _how_ it works - and only for the codebase itself, rather than the behavior of the app.
What would be useful:
- examples of text that can be filtered, and why that would be valuable
- a data flow diagram of runtime behavior, showing how filtering removes unnecessary context
Thanks for your feedback. Will put this in place. Meanwhile, please checkout architecture doc and plugin. The plugin doc could a little bit giving insight of what it does.
I have to agree. I’m interested in the project, so congrats. It’s something I might really like using.
But the one thing I expected to see in the Readme was an example of: takes this tool run output: XXXXXX and converts it to: XX for a savings of 40% of tokens.
This looks like a nice (and useful) project, so thanks for sharing!
I have my own llm wrapping harness, which does this and has a few more tricks. For example, it doesn’t have a lot of mcp but it does have search_mcp and load_mcp tools (and search_skills) so the llm can find what it needs when it needs it without bloating the normal baseline context. The LLMs have proved really good at using them. There is also a waypoint tool they can use to record their thinking in the context without it being the final output. Am thinking about a search_expert to find colleagues it can bring into conversations too. And a lot of other stuff.
Pro tip they worked well for me with response truncation: in the truncated output, say that the full text is available in /tmp/whereever.txt - that way, the llm will be able to query and read more using built in tools without reissuing the big tool call.
great approach. I did that with my opencode based setup as well, it's neat and fun to tune skills and mcp loaders and stuff. Then i got fed up with opencode's design limitations. And then, my own harness work is on hold in favor of a harness-puppeteer paradigm, but that one has also been on hold! I'm mostly currently pulling on the thread of making it easier just to review the voluminous conversation turns!
It has the strip aggressiveness level suport.
You can tune up 3 levels for each template output of your stacktrace using lowfat-filter dsl, shellscript or python.
The LLMs already do that themselves with `tail` all the time. There's a lot of room for improvement on top of that. Though they usually figure it out after a few tries. I often just paste manual runs errors myself anyway.
It requires a bit effort in doing long-term adjustment and tuning for your agent common cli tools commands called. kinda need to evolve on day-to-day basis. But, agent itself can be useful to help tuning this.
Wait, do the coding agents fire `kubectl get -o yaml`??? Most of the harness agents, like CC or codes, are very precise about command construction. For example, the harness add - o and look for the status, for example.
Have terms been established to describe these types of tools? How do I refer to small utilities to perform specific transformations to LLM behavior? CLI filter seems pretty good to describe this tool conversationally but not so much when searching, they some low cardinality keywords.
Great idea. I'm thinking if it could make sense to send the output to a cheap / local model to filter out only the bits that "matter" and pass that through - for the cost some extra time, but maybe it's worth it for saving tokens in the larger model.
I am thinking that a small tool that simply refuses to pass large CLI output to the LLM and warns it to filter the results before reading would achieve this better as the LLM would be forced into thinking and writting the filter itself.
This is a nice little project but I’m weary of sensationally inaccurate titles for stuff like this and the infamous caveman mode. It doesn’t save 91% of tokens: it reduced in one user case 91% of output tokens on the raw CLI output. I am being pedantic about this because these sorts of claims go viral and are inaccurate.
A proper benchmark will compare a large sample of identical prompting with and without the tool, against a specific harness. Once you apply Amdahl’s law, there is no way this saves 91% of tokens holistically, which the title implies.
I work in a non-tech company and these sorts of things keep going viral, with no understanding and with no comprehension of what is actually going on. Engineering is gone and cargo cult magical incantations are in.
Understood. Didn't mean as a click-bait or something. Just sharing my cli report summarize.
Target user here in HN should be tech-savy and this tool is not designed for non-tech because it is required highly customized from user to get the result user want.
Anway, would you mind putting the correct title here ? I will consider to update.
Tools that remove the fat seem like a good idea, but I’m highly suspicious of their effect on the LLM’s reasoning.
LLMs were trained in the typical full-fat output found everywhere on the internet, and all of sudden they get a slightly different response that may look like nothing they have seen before.
Still learning myself, but I've seen MCP tools just lightly wrap upstream json-body REST APIs. Works. But not only is the json structure more tokens but often the model just needs a small subset of fields in the payload.
To be safe if you need a full json, would make conditonal passthrough as the original raw output. Or, need to handle selective object using python via the filter plugin.
He says he adds an output message, but I've tried this myself and I find that quite a lot of the time the agent prefers its own internal monologue over the output of a command.
the bigger problem is agents defaulting to the broadest command possible. kubectl get -o yaml when a jsonpath query would give 1/50th the tokens. filtering after the fact works, but you're still paying for the round trip. better to teach the agent to ask narrow questions in the first place.
Understood. Let me give some examples, most of the time we don't need spaces between table output, git diff produce bunch of unnessary info we just need filename and actual diff lines, kubectl describe we would mostly check for events, image etc etc.
This is the reason why I make it as composable filters as it very depends on your specific ops to optimize the token.
It's a massive red flag to me when you could get decent data to see if your thing actually works, and they don't even attempt to...
Have the LLM use your tool, run it on several of the coding benchmarks. If you're stingy, run it on the ones that don't cost much.
Otherwise, I'm going to assume it doesn't actually work. If it did - Claude, Antigravity, Codex, Pi, or some major player would bundle tools like this into the CLI / harness.
AFAIK, none of the major players do. That's a sign to me these don't work in general.
I've tried building some tools specific to bug fixing. Intelligently feeding context massively helps smaller models. But, what I've found - surprisingly - is that a smaller, much better focused, including a lot of helpful data as well, has almost no impact on larger models compared to what they do by default.
You do save some tokens, though, which is what they're claiming - but not ~99%...
None of the major players are incentivized to care about this, especially not over other opportunities. Why would you expect them to integrate it?
One of the biggest wins you can institute for your own codebase if you use agents is writing your own harness, by a huge margin. The defaults are fine, but you can do better.
Why can I do better than Pi?
I don't want to build my own harness and deal with the bugs... I want to build my project...
My understanding is that Codex / Claude / Gemini subscriptions don't work with custom harnesses.
It's pretty hard to beat 5x more usage if you have the $200/mo subscription by using the API instead.
Your suggestion to using coding benchmarks doesn't really capture the whole picture. I haven't seen a benchmark using kubectl.
> AFAIK, none of the major players do. That's a sign to me these don't work in general.
It's a lose/lose for major players. If it works well, it will lower their revenue. Also there's a high risk it'll significantly worsen results for some people, even if it improves results for others.
There is definitely tons of value to extract from this line of thinking.
VS Code launched it as a feature in their bundled AI functionality last month: https://code.visualstudio.com/updates/v1_121
Defaults imply working...
Reducing tokens and also turns is quite worthless if the LLM doesn’t solve what you put it to do.
Unless something is like 25%+ more cost effective on Gemini for a task, I would not assume those savings are going to transfer to GPT.
If you need to run a test this expensive and slow for every release, hobbiests aren't going to do it.
And if you wanted any broadly specific improvements to coding like they all claim, the costs would be in the thousands per release even for a single for a single model.
And they almost certainly would not be eye popping.
If the models could be SUBSTANTIALLY better, Google and Anthropic and OpenAI wouldn't be finding that out from a hobbiest making wildly unscientific claims.
On the previous large benchmark run, i proved 40-50% cost reduction per correct answer.
I'm not sure why the vendors aren't using token filtering/compression more in their tooling, but perhaps they don't mind users feeding them more data and using more data.
https://github.com/jahala/tilth/tree/main/benchmark
I don't know about cost saving, but if it's keeping the context size down I've had a lot better results using subagents to keep a higher order conversation clean for longer.
What would be useful:
But the one thing I expected to see in the Readme was an example of: takes this tool run output: XXXXXX and converts it to: XX for a savings of 40% of tokens.
This looks like a nice (and useful) project, so thanks for sharing!
Pro tip they worked well for me with response truncation: in the truncated output, say that the full text is available in /tmp/whereever.txt - that way, the llm will be able to query and read more using built in tools without reissuing the big tool call.
A proper benchmark will compare a large sample of identical prompting with and without the tool, against a specific harness. Once you apply Amdahl’s law, there is no way this saves 91% of tokens holistically, which the title implies.
I work in a non-tech company and these sorts of things keep going viral, with no understanding and with no comprehension of what is actually going on. Engineering is gone and cargo cult magical incantations are in.
Target user here in HN should be tech-savy and this tool is not designed for non-tech because it is required highly customized from user to get the result user want.
Anway, would you mind putting the correct title here ? I will consider to update.
But it'll be a case of measuring first, then perhaps a staged integration of a tool like this.
LLMs were trained in the typical full-fat output found everywhere on the internet, and all of sudden they get a slightly different response that may look like nothing they have seen before.
Does that really save tokens in the long run?
[0] https://thetokencompany.com
Harness: I'm about to commit. Good use case Harness: What has changed from X to Y. Bad use case NO?