A Claude Code slash command that estimates what the current session cost at API list prices. It breaks the cost down by model and token type (input, cache writes, cache reads, output, web searches), separates subagent spend, and lists the most expensive prompts.
The built-in /cost shows a total. /costs shows where the money went.
mkdir -p ~/.claude/commands ~/.claude/scripts
An agent workflow worked through a project's issue tracker. Each issue went through three steps, and each step ran in a fresh subagent:
The change: the Implement step moved from Sonnet 5.5 at medium effort to Opus 5.5 at low effort, stepping up to medium only if an issue bounced. Plan and Review stayed as they were.
A Pi cannot build your app. Xcode, xcodebuild, Simulator, codesign and notarytool are macOS-only, and Apple's licence confines macOS to Apple hardware. There is no ARM-Linux workaround and anyone selling you one is wrong.
That is not the interesting question though. Most of what an always-on Mac does for an Apple developer is not compilation — it's being awake. Your laptop closes and travels. A $80 Pi on your home network, reachable from anywhere over Tailscale, is awake for a tenth the price and a twentieth the power of a Mac mini. This is a list of what it can actually do, with the parts that don't work stated plainly so you don't waste a weekend finding out.
Everything below assumes a Pi 5 on a home network, SSH over Tailscale, no ports forwarded on your router. My own is a 16 GB Pi 5 with a 3.6 TB USB disk; most of this works on a 4 GB with an SD card.
Running XCUITest on your own Mac is a bad idea, and the reason is specific rather than aesthetic. testmanagerd bootstraps XCTAutomationSupport into every running GUI app it can enumerate through Accessibility, not just the app under test. A UI test run from one project can therefore reach into an unrelated app you have open and crash it. I lost 38 live terminal sessions that way — a UI test run from project A segfaulted project B's app, which happened to be hosting them:
EXC_BAD_ACCESS in -[XCTAutomationSession initWithAccessibilityFramework:dataSource:]
The fix is a hard boundary: the tests run inside a macOS guest with its own WindowServer, its own apps, and its own TCC database. On the host the only new processes are tart and Virtualization.framework's VM service, neither of which is an app a test runner can attach to. That crash class cannot cross the boundary.
Clone every repository in a GitHub organization, and keep them up to date.
gh), authenticated with gh auth loginbrew install git-lfs && git lfs install)
Serve the current directory over HTTP for local testing, with a live view of the requests coming in.
It is python3 -m http.server with a face on it: the same static file serving,
but instead of a wall of access-log lines scrolling past, you get a fixed panel
showing where it is serving from, a clickable URL, a running request count, and
the last dozen requests with their status codes.
AI coding agents such as Claude Code, OpenCode, Codex, and local-model agents frequently use command-line tools to inspect repositories, search code, process structured data, run tests, and understand changes.
This guide installs a useful set of modern CLI tools on macOS using Homebrew.
A small Zsh script for running Claude Code with a local model hosted by LM Studio.
The script automatically:
