awto-pi-lot: pay-as-you go, accountless mode, cap to infinite loops, etc

[knocte]

Hello folks! 👋

I created awto-pi-lot just for my own needs, but I thought that why not publish it in case it’s useful to someone?

It is an extension that just adds the PPQ.ai provider and its models, especially the most interesting one: AutoClaw.

Why is this useful? For many reasons:

• It mimics same approach as OpenCode Zen, a great provider that allows a pay-as-you-go style in case you’re wary of spending too much via cards/subscriptions.
• It has more models available than OpenCode Zen, such as Grok & DeepSeek.
• It allows a privacy-first approach, as it lets you use AI without linking your data to an email address or billing name (an accountless mode, so to speak).
• Using this provider would also workaround the possible issue that some ToS flag about using third-party software, as you’re not a direct subscriber to the service anymore (no fear of account ban anymore?). This comes up often in pi’s discussions, e.g. Safe to use with Claude/Gemini subscription? #1510 , Question regarding Anthropic's consumer terms of service #1999 , Which model to use with pi-mono? #1945 , etc
• You'll have access to the last frontier models without having to update your Pi: why? you may ask. Because model-provider metadata lives in Pi's source code, so if a new model is released you normally have to wait for a new Pi release that supports it. But not with this extension! Because this extension loads new models dynamically everytime you start Pi, and the PPQ.ai folks are quite fast adopting the new models once they come out.
• Last but not least, PPQ.ai supports a new meta model called “AutoClaw” (unfortunate name, I know, since it can be used anywhere, e.g. with pi, not just with OpenClaw), which redirects your prompt automatically to the most convenient model, in order to be cheaper (in case prompt is simple, it can be redirected to less smart models) and more efficient (faster if the query is simple, as cheaper models will potentially also be faster). More info here: https://ppq.ai/blog/using-autoclaw-with-payperq This becomes even more important for pi, IMHO, because this coding-agent doesn’t have sub-agents support out-of-the-box AFAIK (the point of sub-agent workflows is to not use beefy models with simple instructions, to avoid overspending for simple tasks/prompts/requests).

It can be installed via ./pi install npm:awto-pi-lot. And the code is in this GitHub repo:

https://github.com/tarsgate/awto-pi-lot

And in case you’re wondering why it has so many contributors but only 2 commits, it’s because this used to be a fork of pi-mono instead of an extension and I kept the history of that in the ‘fork’ branch.

The last version is 2.61.0 which signals that I tested it successfully with pi 0.61.0.

PS: Another nice advantage of this pay-as-you-go model that I use myself is to prevent infinite loops on the agent by imposing a cap on its spending: just top-up your account with, let’s say, 3 USD, and let it run. It will have to stop at least until it runs out of funds. No billing surprises this way.

[knocte]

FYI I just released v2.67.68 of this extension. Highlights:

• Fixed compatibility with latest versions of Pi: with v.0.67.68 release there was an error when installing.
• It now correctly filters out models with no tools support (and informs the user about total count of models vs count of compatible models).
• Whitelisted the meta-model auto which at first glance seems to be better than the AutoClaw one, given that the latter seems to be more suited to the "PA" role of an AI, but what we need for Pi is models that help us coding. Considering making this meta-model to be the default (instead of AutoClaw) for next release. (Feedback welcome!)

[knocte]

Version 2.70.2 released some minutes ago:

• Tested with Pi v0.70.2, but there's no reason why it wouldn't work with newer versions.
• If you install from github instead of npm, the download should be much faster because I got rid of the old 'fork' branch.
• Placed meta-model auto as 2nd place in the list after 'AutoClaw', so that it's much easier to find when you type "/model ".
• Fixed potential misleading errors when loading models (e.g. if PPQ.ai is giving weird 4xx errors, the error message now will be clear about this, so that you know that neither your internet connection nor PPQ servers are down).

[knocte]

Version 2.70.6 released yesterday:

• Most important change in this release is how it was published: enabling NPM provenance, so that you have a bit more peace of mind consuming the extension (this way you know that the binary was generated by GitHubActions from a specific commit of the repository).
• Reduced dependencies, so the installation of the extension should be much faster now (regardless if you install from Git or NPM).

However, slight word of caution; even if this release is tagged as 2.70.6 (which means I have tested it with Pi 0.70.6), I detected a regression in Pi 0.70.6 if you consume it from a tarball, so I recommend you stick to 0.70.5 until the fix for it (potentially in PR3998) is merged and released.

[knocte]

Released 2.75.5:

• Tested with Pi 0.75.5
• Pinned dependencies, for security (e.g. supply chain)
• Put back AutoClaw as default instead of Auto (I had put the latter as default in an earlier version, but it's too expensive; will probably develop support for AutoRouter meta-model soon so that you can build your own ladder of models)