PAI-on-Goose (or: Is PAI aimed solely towards Claude Code and Anthropic's Ecology? Intentionally?)

[emory]

Is there anyone else curious about building PAI, based on this home repository, using opensource tools and inference?

As I write about my PAI and DA (Sundae) more, the most frequently asked question I get is "how could I setup something like Sundae for (my kids|project|club)?" and the things that immediately came to mind were privacy and cost, followed by competency and skills required to take care of and feed these things. Setting aside the skill floor for keeping things running, what is the path using opensource software and opensource models?

Every time I imagine having PAI on an stack like that, it's using Goose. I've used Goose a bit, hardly an expert, though I'm not a Claude Desktop|Code expert either. Goose's UI is similar to a Claude Desktop, similar operation too.

Personally I would like this project to be something fairly agnostic and more accessible.

This was a less complicated proposition a couple of months ago, but adopting Skills (not a complaint!) and changes to the context system (also not a complaint!) have underscored that without changing the goals or roadmap of PAI that this will not likely be a "drop-in" replacement for Claude Code and Anthropic.

To determine level of effort and get an idea of how involved it would be now to adapt the PAI upstream (mindful of the further changes post-0.6.0 that are coming in 1.2.0!) and get a better idea of all the components I didn't understand, I asked haiku-4-5 and kimi-k2:1t-cloud about it and put together a small repository and some documentation that was generated.

On that note, here's where I've collected some generated notes on implementing PAI on Goose, at least as far as Anthropic Claude Haiku and Kimi-K2 think it would entail.

github.com/emory/pai-on-goose/

Personally, I am astounded daily with what Sundae does for me. It's been the most exciting computer usage I've had since the first release of Mac OS X. I spend much of my day in terminal windows anyway or Happy, my primary workstation is an iPad Pro.

After reviewing the notes here, "PAI-on-Goose" would probably be best built relying on the unified filesystem context system and other pre-0.6.0 norms and conventions. It's possible Goose will adopt or implement something like Anthropic's Skills, or other ways to extend the capabilities of Goose into other systems, but unless there's a lot of interest in making this project less tethered to Claude, I would probably make a hard fork and base it on the OG PAI repo prior to 0.6.0. Sharing here in case anyone else has had similar thoughts.

[smolcompute]

Does Goose have skills?

[emory]

I am so annoyed I lost a drafted reply so here's an abrupt version - Goose doesn't have Skills. It could, and might at some point but that is the biggest gap identified. I suspect that software like Claude Desktop e.g. Witsy, Flowdown, and Goose and others are trying to figure out how to do that and not if they should. Goose recipes have some overlap though in terms of what is possible.

Skills are currently a vendor lock-in. One that I'm fine with today but likely wouldn't be in the future, but I don't have cycles to steer any implementation work like that. There is a more appropriate thread for the Roadmap and What Should PAI Be and I can elaborate a little on this there but I think many people would also rather have that degree of freedom to choose their interfaces and models that we should probably treat as a commodity considering here among friends I think we're all going to have our preferences but I think we collectively know the model matters less than the prompting or operator?

If I were writing Chatbox or Flowdown I would want to be able to let people drag and drop Skills, and Anthropic hasn't made many fumbles on this sort of thing, mcp is like a year old and has been added to thousands of productivity tools and other software in that time. People have figured out how to add translation layers and tunnel stdio servers over https to mobile clients. I don't see as much agent-to-agent scaffolding but something like Skills is impossible to imagine not wanting that sort of capability to augment and change the way we use software.

[danielmiessler]

I think the concept of skills is extremely universal because it's just marked down in references.

The only part of Skills that is truly Claude Code is the fact that the frontmatter portions are loaded into the system prompt on system load.

But overall commands, workflows, skills, agents, hooks, those are all things that are extremely universal, which is what makes this whole space extremely exciting regardless of what platform you are on.

[emory]

Thanks for the feedback daniel!

re: Skills I agree, but I'm not seeing it get adopted in the tools I use (yet) it's still early but I hope to see them get broader adoption. I don't track the development of any other software like PAI but everyone seems to appreciate mcp enough that there's a thousand marketplaces full of mcp servers and really cool tools like mcpm that let you publish mcp to mobile clients, etc.

I spend most of my time with LLMs in:

• claudeCode (via Happy usually) with Sundae, who is really leaning into her TARDIS era lately "I'm bigger on the inside! I have companions!")
• claudeDesktop (my DA creates a 'mobile aura' for other AIs and applications to use)
• gemini-cli and gemini.app iOS/iPadOS
• tmuxai/aider
• Flowdown
• Raycast
• handful of others (chatbox, claudeSquad, opencode, witsy, ai+ (setapp), aionui - i try out anything really)

[smolcompute]

Honestly, I think Codex has better models (and far more generous limits) but it's very far behind in terms of features compared to Claude Code. I really hope other CLIs catch up on features so we can have a more even playing field and more competition.