Gap Between Vision and Reality: PAI Installation Experience

[Kukuschi]

Gap Between Vision and Reality: PAI Installation Experience

I'm writing this rant as someone who partially fits the target audience Daniel described. I'm an artist and creative first, with moderate programming experience (familiar with several languages, but not a developer by profession). I can navigate technical hurdles better than the average non-technical person, but I still found the installation process challenging.

I wanted to share my experience because it imho highlights a significant tension between the project's stated goals and its current implementation.

From a recent YouTube interview with Daniel Miessler (https://www.youtube.com/watch?v=DbNUDMcEjzY, ~46:00), he described the vision for PAI:

"And at that point, the AI spins up with all the scaffolding to help you with meal planning, to help you with like encouraging you, to help you find other artists, right? Because this I'm not trying to build a product for tech people. Tech people are already techie, right? This is not about coding. This is about enabling a human to be better at what it is that they want to do to help them activate their full self."

In another video (https://www.youtube.com/watch?v=Le0DLrn7ta0 ~19:12), Daniel shares his love for the command line:

"Terminal is my favorite place to live. And I love the fact that there's documentation, there's flags, there's switches, there's options... And it means you know how to use it."

• while presenting a graphic that presents GUIs as problematic.

My Installation Journey (~ 3h)

Started on Windows ... did not work
Moved to WSL (Windows Subsystem for Linux) - a non-technical person would likely have given up here already
What followed was a series of manual steps that weren't obvious to me (because "The installer will: Detect your system and install prerequisites (Bun, Git, Claude Code)" did not do any of that for me):

• Node.js installation (manual)
• Claude CLI installation (the quickstart script at https://code.claude.com/docs/en/quickstart - "curl -fsSL https://claude.ai/install.sh | bash" - did not work)
• zsh installation (to replace bash)
• unzip installation
• Bun installation
• Manual PATH variable configuration

After multiple hurdles, PAI finally installed (with some errors/warnings) and runs. But using it isn't exactly comfortable for me.

The Gap

Stated Goal: "Not for tech people... not about coding... enabling a human to be better at what they want to do"
Current Reality: The installation and usage experience requires:

• Terminal knowledge
• Comfort with manual configuration
• Willingness to troubleshoot failing scripts
• Knowledge of what tools need to be installed

For someone who doesn't already "live in the Terminal," each of these steps is a potential drop-off point.

I'm sharing this because I genuinely believe in (most of) the vision of PAI. The installation challenges aren't insurmountable for me, but they do highlight a significant barrier to entry that seems inconsistent with the stated mission of being "for everyone, not just techies."

[realGregorVucajnk]

I can understand your frustration. What I can offer is a perspective:

• this space is developing so rapidly, where keeping up to date is nearly impossible but possible with TUI
• while learning curve of terminal can is steep, it's worth it. No better way to learn is to just use it. However, it's not all hardcore learning:
• • use terminal solutions like WARP and just ask AI to help you fill the gaps. It's still the terminal and you can use it as a terminal or talking to AI to use the terminal.
• • use Claude CLI to help you install and update. I've not touched the upgrade path since pre-2x version, I just work in terminal with claude and "talking to it" to upgrade (modify, migrate, test, verify, fix..) my PAI installation. PAI is actually really good in self-fixing.
• once you have a fixed upgrade path, then create your own script / TUI or GUI workflow.

Hope this helps.

[garthsch]

+1 on the friction outside the macOS ecosystem. I’m seeing frequent silent failures in WSL and Linux where PAI assumes it’s running on a Darwin-based system.

While I’m personally enjoying the backend deep-dive, we have to be honest about the current architecture: PAI is a Mac-first tool. It offers a smooth path for users with high-end hardware and the budget for Claude Code/API credits, but that’s a high bar. To put it in perspective, someone who can afford a Mac and a US$20/month subscription is likely in the top 1% of global earners.

If the mission is "AI should magnify everyone—not just the top 1%," we can’t have an entry price that defaults to a Silicon Valley-standard tech stack. We can’t expect a student in Southeast Asia on a second-hand Chromebook to use this for "human activation" if the system silently fails on their OS.

As a starting point, I’d recommend a stricter logging or feedback system. I only caught these failures because I was digging into the hook structures and noticed my WSL environment wasn't triggering them. PAI is excellent at repairing issues, but it needs the observability to see them in the first place. If the system doesn't perceive the friction, it can't automate the fix.

[jarednwilson]

Hey man, your experience was exactly like my installation experience with PAI, but I can tell you it has all been worth the pain. This system truly is the frontier, and it can do practically anything you can imagine, so it's definitely worth it.

The way I cut the learning curve was I basically took screenshots of what error message I was having, and just kept giving it to ChatGPT, and it kept telling me what to do to fix it. I just bumbled my way into getting Claude code installed on Windows. At the time I didn't know that it was built for Mac, so I had the PAI system self-heal itself for Windows. It basically just rewrote the whole thing into a custom Windows version, and that was very painful, but the PAI system did it. Now I'm using WSL Linux, so it's a little bit easier, but I'll never whole-hog install a new version; I only cherry-pick new features into my existing PAI because apple PAI never one-on-one works in Linux.

[Kukuschi]

The point I was trying to make wasn't about my personal frustration (i got it running), it was about perspective.

Tech people have a tendency to underestimate how much friction the tools they feel at home in generate for non-techies (i cal it the Vim-Syndrome ^^). The responses here actually prove my point:

"Use terminal solutions like WARP and just ask AI to help you fill the gaps"
"I just bumbled my way into getting it installed"
"It has all been worth the pain"

To someone who doesn't live in the terminal, "just ask AI to fill the gaps" and "bumble your way through it" describe a fundamentally different relationship with computers. For many people, the terminal isn't a space they visit, it's a space they actively avoid because it feels like staring at a blank page in a foreign language.

I agree with @garthsch's observation about PAI being Mac-first. But beyond the macOS/Windows divide, there's a deeper question:

If the installation experience requires "bumbling through" and accepting "pain" as the price of entry even for people with some tech skills, is it really meant to be a product for "non-tech people"?

The "it's worth it" argument only works for people who can see the destination from the starting line. For someone who doesn't yet know what PAI can do, each terminal error isn't a stepping stone, it is a wall.

I'm not asking for personal hand-holding. I'm asking whether the project's stated mission and its actual onboarding experience are aligned. Right now, imho they're not.

P.S.: Concerning the $20/month subscription: I haven't tried it yet, but from what I've heard, you'd run out of tokens very quickly with that plan. Since $20 is already too much for me, I needed a cheaper solution and stumbled across this (not affiliated). It works quite well so far (though I can't compare it to Claude's models directly).

[Drizzt321]

Environment: Debian/Stable (13.3) linux, desktop & laptop & Android (LineageOS) phone

So my recent experience, mostly smooth, although the installer went X11 compat, rather than native Wayland (IMO @danielmiessler this should be a low priority TODO, Linux world is almost entirely gone Wayland at this point). The single worst/annoying part is that I don't have any voice in/out setup like you do (or currently want, might want to in the future). So those setup pieces were a bit annoying and yet my PAI (named Metis) still insists, often enough, on doing a bashtool callout to try and provide the voice output. A very simple boolean setting/flag someone would be great.

Things that did go well, is I use SyncThing, and just pointing SyncThing at my ~/.claude/ directory with the following ignore:

// Machine-local per-session caches — regenerate at session start, no sync value
MEMORY/STATE/model-cache.txt
MEMORY/STATE/session-name-cache.sh
MEMORY/STATE/events.jsonl

And prompting, as part of my early interactions with Metis, to have it be aware it might be on my desktop vs laptop, and general capabilities of each (e.g. laptop might be offline sometimes). Haven't switched back and forth on that yet, but intended too and if I report any real issues, will create some bug tickets as needed. This aspect is important to me, as I'll to have similar expectations of interaction and capabilities between where I run PAI, while also having "I know where I'm running, what that means for how I can interact with the world".

I think I'd like to slightly pushback on some of the others, I think while it's slightly MacOS centric, I think even more so it's Unix-based system centric. I've had extremely little friction (except for my own specific "want to modify things" stuff) with the setup, and terminal is a native interface for me long before this.

I agree this is a stumbling block for a lot of somewhat less technically savvy people, and most especially on Windows. Unfortunately...Windows is and probably always will be a 2nd class *nix citizen. WSL2 was big progress, but fundamentally it's a well integrated fancified VM/emulation of a Linux system, rather than it being fully native Windows shell system. Since I don't run Windows anymore, I can't really help provide fixes/testing for this.

[Kukuschi]

With respect, this:

"I've had extremely little friction... terminal is a native interface for me long before this"

is exactly the perspective I was critiquing in my original post.

You're describing why the installation works for you, someone who is at home in the terminal, comfortable with Unix systems, and for whom SyncThing, Wayland differences, and session cache ignore patterns are a natural part of the setup.

That's completely valid but also completely misses the point and does not address the question I raised:

Is this tool meant for people who don't think that way?

The project vision talks about "enabling humans to activate their full self"—not "enabling humans who already know how to configure session cache ignores."

PAI's mission is twofold:
Activate as many people as possible — Help people identify, articulate, and pursue their own purpose in life through AI-augmented self-discovery
Make the best AI available in the world accessible to everyone — Ensure this quality of AI infrastructure isn't reserved for just the rich or technical elite.

The observation that Windows is a "second-class *nix citizen" may be technically accurate. But a non-technical person sitting at a Windows machine (and there are many of them) hears: "This tool isn't for you." even though the project's stated mission promises the opposite.
This isn't about Windows-bashing or whether WSL2 is "good enough." It's about the core question:

If this tool is supposedly "not (only) for tech people", why does the entry point assume you're already at home in the techies' universe?

I'm genuinely glad Metis is running smoothly for you. But your response unfortunately confirms the very problem I described:
those for whom it works don't see why it doesn't work for others.

[Drizzt321]

Fair enough, although the SyncThing/etc is definitely FAR beyond what should be a more regular user experience, or at least some future "here's some simpler ways to sync around" should be developed. Longer term IMO, given how in flux PAI is still.

Good philosophical/intention, you're right. MacOS might be rather popular, but that also generally denotes a set of people that will tend to already have more opportunities/money/etc. Where this would be amazing is those who have more difficulty, or lower end cheaper old Windows hardware. Or, hell, a Chromebook type device, although I'd see difficulties getting this on a Chromebook.

I think it'd also be a valuable experiment for Daniel or some of us to use this exclusively with a standard Pro individual plan, where usage/tokens are a lot more precious. I personally rapidly when from Pro to Max, only took me 1-1.5 weeks to decide to do that. Before I was using PAI even.

I don't think, didn't intend, for my response to be "I don't see a problem with why it doesn't work for others". Wasn't meant to at least. I think the 2nd part was more about "why maybe it doesn't work well on Windows". Which isn't a help necessarily, that's true.

But yes, I think you are right to point out the failure of installing and running moderately smoothly (this is still early days, some friction IS to be expected) on Windows to work towards the core principle of "for everyone".

Although thinking about it, I wonder if it's possible to do Claude Code via Desktop application, point it to a clone of this repo, and say "look at the installer, and does what it will do for the X version into your .claude/ directory". Or some variation of that.

[AojdevStudio]

Right now the reality is that anything AI is built with windows users as an after-thought. It's the truth, even the big labs, their first release is always on MacOS and then in the fine print they say windows version coming soon. Sadly that's where we are, even though i'm not a windows user i was inspired to try an ease the friction as much as humanly possible to get a windows machine ready for the Agentic world in 2026. I have an open-source project called dev-bootstrap it's one command ran in the terminal the initial premise is to help the non-techy onboard to the world of AI.. but it's seeming that there's more to it than that. I'm inspired to find a happy medium. PAI should be available to all. Instead of waiting for daniel to do something about it, we should all just hack at it with our PAI's.

[MikeRouten]

I find this conversation very .interesting as I'm asking myself where I want to install my on PAI? I'm currently doing my day-to-day work on a windows laptop. I have an older PC tower that has Linux Mint loaded on it that I don't fire up very often and I have zero Mac products.

Has anyone hosted a PAI in a VPS on a cloud hosting service? I've heard that Hostinger has native Open Claw installs, so that got me wondering if it would make sense to install PAI on a cloud based service to overcome the limitations of running on local hardware.

I think I'm going to start by clearing off that old tower PC and try and get it running locally first.

[zeddy303]

I'm running PAI on Hostinger VPS and it works great. I basically have a tmux shell all the time, either working on my phone while taking the train, connecting while at work on my laptop. It keeps the session intact when moving between 5G, no signal, elevator and wifi. I have a passwordless set up and use tailscale to connect to it while blocking everything else. Then i can have PAI set up docker images for additional capabilities all in one location that's not messing aroundn with my local machine.

[MikeRouten]

Thanks for the detailed reply.