Apple is suing OpenAI for trade secrets theft. The short version from TechCrunch: job candidates were allegedly asked to bring unreleased Apple hardware components into their OpenAI interviews. One employee reportedly joked about unauthorized system access. The Verge has the six wildest claims if you want to read it slowly. I do not know what outcome Apple wants here. But I know what this signals for operators: the major labs are competing so hard right now that the floor is coming off the norms, and that changes the risk surface of everything downstream.

The more immediately useful story is in the repos today. OpenClaude-Portable hit over a thousand stars doing one thing: it lets you run Claude Code from a USB drive on any PC, no installation. I tested the concept mentally and the use case is real. Air-gapped machines, locked-down corporate laptops, client sites where you cannot install software. One USB drive, full agent environment, leave no trace. That is the kind of lateral operator move that does not get written up in a keynote.

Nine repos passed the bar this week. Lead with the one that changes where you can work.

New here? This part is for you. Operators, skip ahead.

What is open source? Open source means the code is public and free for anyone to read, use, and build on. The nine repos below are all open source. That is why you can just grab them: no purchase, no permission. Someone built a useful thing and chose to give it away. A huge part of the AI tooling world runs on exactly this.

Do this today (2 minutes, no coding): Open GitHub in your browser and look at any one of today's repos. Read the top of the page, the README, the project's plain-language note on what it does. You are not installing anything. You are learning to read a project page, which is the first skill for using any of this.

How to use today's picks: Today is nine repos plus a way to run Claude from a USB drive. You do not need to try them all. Pick the one whose one-line pitch sounds like a problem you actually have, click through, and skim its README. That is the whole move: pitch, click, skim. Installing comes later, only if it earns it.

Plain English. README: the front page of a code project, with plain-language notes on what it does and how to start. Always read this first. local: running something on your own computer instead of on someone else's server in the cloud. Claude from a USB drive is as local as it gets.

We read your replies. You told us you are here to learn, so we built this for you. Hit reply anytime with a word you want decoded, and we will define it here. That is the whole point: we learn this together.

[Repo] OpenClaude-Portable, Run Claude Code from a USB drive on any PC, no installation required. 1,033 stars. The operator use case is immediate: locked-down corporate machines, client sites, air-gapped environments. You plug in, you work, you leave nothing behind.

[Repo] claude-code-hooks-mastery, A focused repo for mastering Claude Code hooks. 3,826 stars. If you are not using hooks to auto-enforce rules at the event level (lint on save, block a dangerous bash, log every tool call), you are supervising Claude manually, which scales to exactly zero.

[Repo] claude-code-skill-factory, An open-source toolkit for building and deploying production-ready Claude skills, agents, slash commands, and LLM prompts at scale. 829 stars. Build the skill once, call it from every session. The factory structure keeps your skill library from becoming a graveyard of one-off prompts.

[Repo] claude-modular, A production-ready modular Claude Code framework with 30+ commands, token optimization, and MCP server integration. 286 stars. The value is in the command organization system: structured command hierarchies mean Claude spends less of your context window figuring out what you want and more of it doing it.

[Repo] claude-replay, Converts AI coding agent sessions (Claude Code, Cursor, Codex, Gemini, OpenCode) into self-contained, embeddable HTML replays. 750 stars. The operator use case most people miss: show a client exactly what your agent did, step by step, without giving them repo access. Auditable, shareable, no setup required on their end.

[Repo] AutoHedge, A swarm-intelligence agent stack for autonomous market analysis, risk management, and trade execution. 3,800 stars. The architecture is the interesting part: multiple specialized agents coordinating on a shared task, which is the pattern, not the finance use case. Transplant the orchestration structure into your own domain.

[Repo] FinceptTerminal, A modern terminal-based finance application with advanced market analytics, investment research, and economic data tools. 28,337 stars. Not an AI agent, but the data layer that most financial agents are missing. If you are building anything in the finance vertical, this is the data substrate worth knowing.

[Repo] OpenHands, Open-source software development agent with sandboxed execution, web browsing, and code editing. The benchmark comparison if you are evaluating autonomous coding agents: this is the bar that the Claude Code ecosystem is competing against. Know what you are comparing to.

[Repo] deer-flow, An open-source long-horizon SuperAgent harness from Bytedance that researches, codes, and creates with sandboxes, memories, tools, skills, subagents, and a message gateway. 76,932 stars. The architecture is worth reading even if you do not deploy it: this is what a production multi-agent system looks like when a major lab builds it for internal use and then open-sources it.

[MCP] claude_code-gemini-mcp, Drops Gemini directly into Claude Code as an MCP server. 245 stars. The non-obvious use: run Gemini's long context window for the retrieval pass, hand the result back to Claude for the action pass. Two models, each doing what it is actually better at, inside one workflow.

[Tool] Opptrix, An open-source LLM-driven investment research assistant for China A-shares with 40+ MCP tools, factor screening, backtesting, watchlist management, and an Electron desktop app. TypeScript and React monorepo. The architecture is worth a look for any operator building a research or analysis agent: 40+ tools scoped to a single domain is a clean example of how to instrument a vertical agent without losing control of it.

- Open-weight models now handle 29% of production AI traffic. Vercel's AI Gateway Production Index for July 2026 shows price per token has also flattened. The operator consequence: the cost floor argument for closed proprietary models is thinner than it was 90 days ago. If you have not re-evaluated your model routing lately, this is the data that should move you.

- Context bombing is now a defensive technique against hacking agents. Ars Technica reports defenders are deliberately flooding attacking agents with context to exhaust their window before they can act. Prompt injection is now a two-sided weapon. If you are running agents that process untrusted input, this attack surface is now more complicated than it was last month.

- The model price war is compressing costs further. The LA Times covers OpenAI, Meta, and xAI all cutting model costs this week. GPT-5.6 is designed to complete more work on fewer tokens. Grok 4.5 is in the mix. The operator read: your inference budget should be going down this quarter, and if it is not, you are either overusing a premium tier or your routing is broken.

- GLM-5.2, a free open-source model from China, is drawing DeepSeek-level comparisons. Business Insider reports developers and executives have spent a week praising its output quality. The pattern is repeating: another open-weight model closes the gap with frontier labs. The standing question is what you still need a closed API for.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: the handoff test every automated workflow eventually faces.

An automated workflow that only its builder can operate is not automation. It is a bus factor of one wearing an automation costume.

This week we handed a daily AI-run workflow to a non-technical operator. Not a developer, not someone who will read the code. The first draft of the handoff doc explained how the system works: the components, the flow, the why. It was accurate, and it was useless. The operator does not need the architecture. They need to know what today is supposed to look like, and what to do when it does not.

The doc that worked fit on one page and had a completely different shape. It described states, not mechanisms. The system as they will meet it: here is what a normal morning looks like, down to the specific things you should see. Here are the three ways it looks when something is wrong, in plain words, with a screenshot each. For each bad state, exactly one action, do this, or send this exact message to this person. And at the bottom, the most important section: the do-not-touch list, the things that look like buttons a helpful person might press to fix things, and must not.

One page. States, one action per state, an escalation path, a no-touch list.

The pattern generalizes past AI workflows to anything you run: write the runbook for the person who will meet the system at its worst moment with the least context. If your handoff doc explains how it works before it explains what wrong looks like, it is written for the wrong reader.

Could someone who has never seen your system run it for a week from one page? Hit reply and tell us what would break first. We read every one.

Apple suing OpenAI over stolen hardware is the headline. The more durable story is a USB drive that lets you run a full agent environment on a machine you do not own. Nine repos today. The one that changes where you can work is the one worth clicking first.

See you Wednesday.