I got Claude AI running natively on a PowerPC iBook G4 (Tiger)

[watermark_hd]

I got Claude AI running natively on a PowerPC iBook G4. Here's what it can actually do.

Type claude in Terminal, and you get a real coding assistant that runs on the G4 itself — no cloud IDE, no remote desktop, no cheating. It can:

• Hold a conversation, in English or Japanese (it figures out which one you're using automatically)
• Read and write files on your Mac
• Run shell commands for you
• Actually help you write, fix, and explain code, right there in the Tiger terminal

Setup is stupidly simple: double-click an installer, paste an API key, type claude. That's the entire interface. No Xcode, no MacPorts, no command-line archaeology required on your end — I already did that part so you don't have to.

And yes, this is the same Mac whose Photoshop filters crawl, whose marquee selection lags a full beat behind your cursor, and whose fan screams like it's applying for a job at NASA every time you touch a layer mask. Let your M-series Mac handle that work. Your PPC Mac gets a new job instead: coding consultant.

A couple of things to know before you dive in:

• You'll need your own Anthropic API key (console.anthropic.com — separate pay-as-you-go billing, NOT your claude.ai subscription; a few dollars covers a lot of casual use)
• Tested on an iBook G4 (PowerBook6,5) running Tiger 10.4.11. Other PPC Macs on Tiger should work too, but haven't been tested on every model
• Under the hood: Tiger's stock OpenSSL can't even do a modern HTTPS handshake, so this required building a fresh OpenSSL + curl locally on the G4 itself (no MacPorts — that hits its own chicken-and-egg TLS problem), driven by a small dependency-free Perl script. All of that is already built and bundled, you just install it

Download / install instructions : https://github.com/watermark-hd/ppc_claude_cli

Old Macs deserve better than a shelf. Put yours back to work.

v1.1.0 — fixed arrow keys inserting garbage characters instead of moving the cursor (found this using it on the iBook myself). Editing input now works properly, including for Japanese text.

Fix: Long input (and Up/Down history recall) used to duplicate the same line over and over on screen. Fixed the line editor's redraw so it clears the whole wrapped block, not just the last row.

 

[barracuda156]

I will be interested in having Claude Code agent for ppc. That will help me to fix Qt6, for example, and perhaps NodeJS and ocamlopt.

 

[watermark_hd]

I will be interested in having Claude Code agent for ppc. That will help me to fix Qt6, for example, and perhaps NodeJS and ocamlopt.

Yeah, it's not a toy demo — it's built to hit the Claude API directly and run a real agent loop (Claude does the actual thinking/coding, the Perl side is just the four tools + HTTP plumbing). Right now those tools are read_file / write_file / list_dir / run_shell, but run_shell means it can already drive make, gcc, patch, git, etc. directly — so a "read the build error, patch a file, rebuild, repeat" loop should already work in practice, not just toy examples.

Biggest catch was actually just getting it online at all: Tiger's stock OpenSSL (0.9.7l) can't even do TLS 1.0, and Anthropic's API requires 1.2+. Had to cross-build a newer OpenSSL + curl from source, compiled locally on the G4 itself, before the agent could even talk to the API. Also needs your own Anthropic API key (pay-as-you-go, not bundled).

Genuinely curious whether it holds up on something real like a Qt6 build fix — happy to share the setup scripts if you want to try it on one of your PPC boxes.

 

[featherwhisker]

I don't think vibecoded software like this being released is beneficial to anyone, especially if you cant even bring yourself to do as little as writing your own post.

Download / install instructions: [your link here]

Genuinely what the hell is this? Did you just post the direct output from ChatGPT without even skimming through it?

 

[watermark_hd]

I don't think vibecoded software like this being released is beneficial to anyone, especially if you cant even bring yourself to do as little as writing your own post.


Genuinely what the hell is this? Did you just post the direct output from ChatGPT without even skimming through it?

Yeah, I use AI a lot — I make plenty of mistakes without it, and English isn't my first language either, so some of this may read a bit rough. I built this for myself and run it daily on my own iBook; just wanted to share it in case anyone else wanted the same thing. No money or credit in it for me either way.

 

[Doq]

Okay, when I saw the title saying "natively" I thought this was like, loading and running models locally and thought "ok, but why?"

But then I read the OP and then thought "hm, interesting, but why?"

 

[watermark_hd]

Okay, when I saw the title saying "natively" I thought this was like, loading and running models locally and thought "ok, but why?"

But then I read the OP and then thought "hm, interesting, but why?"

Honest answer: on this iBook I couldn't even get past the login screen of any AI chat site, and cropping a photo in Photoshop takes 10 minutes. But I didn't want to send it to the shelf. A browser is too much for a G4 (1.2GHz / 1.25GB here) — a text terminal isn't. The thinking happens in the cloud; the iBook just moves text around, and it's still great at that.

About "natively" — full disclosure, Claude wrote the English for that post. I stopped trusting my own English after an American told me I look "like Woody" and I spent a good minute wondering why he was asking if I liked Toy Story. What I meant: the client itself — OpenSSL, curl, everything — was built and runs on the G4. No proxy machine, no remote desktop.

So the iBook got a career change: coding consultant.

 

[f54da]

>Claude wrote the English for that post. I stopped trusting my own English

Fyi most people will find Claude-speak much more off-putting than broken english. You will be better off (in terms of people not dismissing you) if you type in your native language then use google translate. Otherwise if you can at least cobble together something in English, most english speakers will not really mind even if it's not completely grammatical

 

[watermark_hd]

>Claude wrote the English for that post. I stopped trusting my own English

Fyi most people will find Claude-speak much more off-putting than broken english. You will be better off (in terms of people not dismissing you) if you type in your native language then use google translate. Otherwise if you can at least cobble together something in English, most english speakers will not really mind even if it's not completely grammatical

Thanks for pointing that out. I had been hesitant, wondering if I’d make sense or get laughed at—after all, English and Japanese have completely opposite word orders! Still, I do feel embarrassed about having just pasted the AI-generated text exactly as it was. I’m truly grateful to the people who help me realize these things through their feedback.

 

[f54da]

If you're japanese then you probably already know this, but use DeepL for translation back/forth? It's better than google translate at capturing nuance and technically uses an LLM but optimized for translation accuracy and matching style rather than a "chatbot" that imposes its own voice and persona.