dnlcrl

Inside the Claude Code source

Anthropic's Claude Code CLI source code leaked onto GitHub recently. All of it. About 1,900 files and a lot of TypeScript.

I read through the key modules. What follows is a breakdown of the surprising parts: how the system actually works, where Anthropic made clever engineering choices, and where their approach diverges from OpenAI's Codex in ways you wouldn't guess from using either tool.

Lifecycle of a request

Here's what happens when you type a message into Claude Code:

Raspberry Pi on Internet via reverse SSH tunnel

HackerNews discussed this with many alternative solutions: https://news.ycombinator.com/item?id=24893615

I already have my own domain name: mydomain.com. I wanted to be able to run some webapps on my Raspberry Pi 4B running perpetually at home in headless mode (just needs 5W power and wireless internet). I wanted to be able to access these apps from public Internet. Dynamic DNS wasn't an option because my ISP blocks all incoming traffic. ngrok would work but the free plan is too restrictive.

I bought a cheap 2GB RAM, 20GB disk VM + a 25GB volume on Hetzner for about 4 EUR/month. Hetzner gave me a static IP for it. I haven't purchased a floating IP yet.

Complete solution for websites hosting

This gist contains example of how you can configure nginx reverse-proxy with autmatic container discovery, SSL certificates generation (using Let's Encrypt) and auto updates.

Features:

Automatically detect new containers and reconfigure nginx reverse-proxy
Automatically generate/update SSL certificates for all specified containers.
Watch for new docker images and update them.
Ban bots and hackers who are trying to bruteforce your website or do anything suspicious.

	git branch -m main master
	git push -u origin master
	git remote set-head origin master
	# Now go to github in a browser,
	# open the repo, click settings, branch and change the default branch to master
	git push origin --delete main

	import os
	import tarfile
	import imageio
	import tqdm
	import numpy as np

	def get_nomnist():
	classes = ["A", "B", "C", "D", "E", "F", "G", "H", "I", "J"]
	tar_path = "./data/no-mnist/archive/notMNIST_large.tar"
	tmp_path = "./data/no-mnist/archive/tmp"

	# WARNING: These steps seem to not work anymore!

	#!/bin/bash

	# Purge existign CUDA first
	sudo apt --purge remove "cublas" "cuda"
	sudo apt --purge remove "nvidia*"

	# Install CUDA Toolkit 10
	wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.0.130-1_amd64.deb

	import torch
	import torch.nn as nn

	class Model(nn.Module):
	def __init__(self):
	super(Model, self).__init__()
	self.linear = nn.Linear(1,1)

	def forward(self, x):
	y = self.linear(x)

	"""
	A series of speed tests on pytorch LSTMs.
	- LSTM is fastest (no surprise)
	- When you have to go timestep-by-timestep, LSTMCell is faster than LSTM
	- Iterating using chunks is slightly faster than __iter__ or indexing depending on setup

	Results

	My Ubuntu server:
	OS: posix, pytorch version: 0.4.0a0+67bbf58

	import torch
	import pickle

	ts = [torch.ones(10, 20) + n for n in range(4)]
	diff = 0

	with open('tensor.pickle', 'wb') as f:
	for t in ts:
	start = f.tell()
	torch.save(t, f)