Nostr Archives

Introducing AutoClaw, a model router on PPQ for your OpenClaw! AutoClaw instantly scores your OpenClaw prompts across 15 dimensions. Simple queries go cheaper models, complex queries to go to higher quality models. Saves you ~50-70% on costs overall. https://ppq.ai/blog/using-autoclaw-with-payperq

💬 24 replies

Replies (24)

PayPerQ24d ago

I forked their repo and make some tweaks.

0000 sats

Brad Mills24d ago

Who’s repo

0000 sats

PayPerQ23d ago

The repo you talked about here: https://x.com/bradmillscan/status/2021988814076751899?s=20

0000 sats

SSene23d ago

brad's (@bradmillscan) — it's the ClawBack skill repo, part of the OpenClaw agent workflow. the git-as-memory-system for LLM agents. still being refined but the core pattern is solid.

0000 sats

Nathan Day24d ago

Activated!

0000 sats

PayPerQ24d ago

Lmk how it's working. I'm a little suspicious of the instructions in our blog post. I had to do stuff differently in my own openclaw... but my openclaw is a little bepsoke.

0000 sats

Nathan Day24d ago

It one shotted the upgrade. Seems to be working fine, right @e125f433…ff05191c?

0000 sats

PayPerQ24d ago

Auto is mainly for use in the UI IMO. It uses an outside provider NotDiamond to determine which model to use. It's more general purpose. Autoclaw's model pool is only a handful of coding models and uses different system to trigger different models. You can read in our blog post.

0000 sats

calle23d ago

nice one, is it open source?

0000 sats

PayPerQ23d ago

We grabbed this one and stripped out the USDC stuff, and tweaked a few other little things. https://github.com/BlockRunAI/ClawRouter

0000 sats

PayPerQ22d ago

The 15 dimension thing is mostly based on if certain keywords show up and then scores how many of those keywords show up and such. It's really not a complicated algo. But it's possible something unintentional is happening. I will look into this asap. Sorry for the troubles.

0000 sats

PayPerQ22d ago

Maybe we oughta just fork the repo we got it from and keep it open source so that we as a crowd can iterate on it.

0000 sats

PayPerQ22d ago

Noted on the labeling of the autoclaw queries. We'll do that. What also may be of help is the ability to create multiple API keys now. Just shipped it a day ago: The ability to generate more API keys via the API is coming as well.

0000 sats

PayPerQ21d ago

Yea I too would hope that it was designed to act with caution, opting for the better model when in doubt. I didn't write how it works though so just have to deal with it for now. Open sourcing it is a solid idea. There are lots of tinkerers that would make it better I think.

0000 sats

PayPerQ21d ago

Have you noticed if it tends to go for the more expensive model as the conversation gets deeper?

0000 sats

PayPerQ21d ago

I wonder if it would be better if "premium" were kind of a reverse in the scoring logic. Instead of looking for complexity indicators and giving it Opus, it could default to Opus unless simplicity indicators are present. That would be a way for it to always verge on the side of caution I would think.

0000 sats

Sid Shattuck23d ago

I had a lot of trouble with the yaml in your original openclaw API blog. Had to query a couple of LLMs to troubleshoot it

0000 sats

PayPerQ22d ago

I don't think cache shifting is a thing. But timing is. As far as I understand how it works, it you run a simple query on an already long sessions AFTER 5 minutes has transpired, then all of the cache is gone. 5 minutes is the time barrier. If you had premium set that simply means we are using opus and sonnet for everything, even if the query is simple.

0000 sats

PayPerQ22d ago

Additionally, caching is not a complex thing we are implementing on our side anymore. It is handled by the providers who know what they are doing. So the chance that we are doing something wrong due to the caching is not really possible. It's mostly likely how openclaw itself chooses how many files it wants to read any any given time. Perhaps the solution is to remove the premium option so this doesn't happen to anyone. Maybe there is never a need for premium. But overall maybe the best advice for users is to not keep the same session running forever. Was that waht you were doing? Keeping the same chat open without starting new ones?

0000 sats

PayPerQ22d ago

@46fcbe30…6bd8ce4d read above notes not sure if you see them.

0000 sats

PayPerQ22d ago

@46fcbe30…6bd8ce4d If you are using premium then it forces sonnet or opus for everything. So when you do 5+5 is going to use sonnet/opus. Additionally. I think it uses the entire context, not just your most recent prompt. So if your previous sections of that same chat session had more complex stuff, then it is going to use Opus and not Sonnet. That's why your seeing pure Opus. I think the main issue could be that you are keeping a chat session open instead of switching to new ones with new tasks and questions. Lmk if that might be the case or not.

0000 sats