We read every reply you send us. Lately a lot of them say the same thing: you are here to learn this, not just watch it scroll by.

So we listened. Starting today, Start Here leads every issue: a plain-English on-ramp for anyone on day one. No jargon, no terminal required. The AIgent was a community before it was a newsletter, and the front door should show it.

Operators, your depth is still here, one scroll down. Beginners, you are exactly who we built this for.

Below: nine repos worth cloning, three tools, and four signals. Let us get into it.

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

What is an AI agent? A chatbot answers, then waits. An agent takes actions on its own. You give it a goal and it works the steps to get there: reading files, running tools, checking its own work. Picture the gap between a calculator and an intern. One responds. The other does the task.

Do this today (2 minutes, no coding): Open Claude or ChatGPT and give it a real errand, not a question. Try: 'Here are my 5 tasks this week. Plan them and tell me what to do first.' Watch it reason through the steps instead of just replying. That shift is the whole game.

How to use today's picks: The Drops and Stack just below are tools builders use right now. You do not need to install anything to get value today. Read the one-line pitch, click through, and skim the README, the project's front-page note on what it does and how to start. You are learning the landscape. Installing comes later, when one of them solves a problem you actually have.

Plain English. repo (repository): a project folder of code and files, stored online, usually on GitHub. 'Clone a repo' just means download a copy. MCP (Model Context Protocol): a standard plug that lets an AI use outside tools and data, so it can act, not just talk.

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] affaan-m/ECC, A drop-in agent harness for Claude Code, Cursor, and Codex. Ships with 63 agents, 249 skills, persistent memory, security scanning, and automation hooks. If you are building multi-agent pipelines and wiring your own memory layer from scratch, start here instead.

[Repo] colbymchenry/codegraph, Pre-indexed code knowledge graph that auto-syncs on every code change. Supports Claude Code, Codex, Gemini, Cursor, and several others. Fewer tokens per tool call is the whole game at scale, and this is the cleanest implementation I have seen surfaced from community this week (59,409 stars).

[Repo] Dicklesworthstone/destructive_command_guard, Blocks dangerous git and shell commands before an agent executes them. 2,708 stars and trending today. This is the repo I install on every new agent project now. The gotcha it prevents: an agent with write access to a production shell doing exactly what you told it to, wrong.

[Repo] wonderwhy-er/DesktopCommanderMCP, MCP server that gives Claude terminal control, filesystem search, and diff-based file editing. 7,944 stars. The diff editing is the part worth the install: it patches files surgically instead of rewriting them whole.

[Repo] future-agi/future-agi, End-to-end platform for evaluating and observing LLM and agent applications. Tracing, evals, simulations, datasets, guardrails, and a gateway, all self-hostable. If you are running agents in production without an eval layer, this is the cleanest open-source option I have found (1,382 stars).

[Repo] openai/whisper, solid speech recognition via large-scale weak supervision. 104,810 stars. The reason it keeps surfacing: it is still the most reliable open-weight transcription baseline for any pipeline that takes audio input. Pipes cleanly into agent workflows.

[Repo] langgenius/dify, Open-source LLM app platform with drag-and-drop RAG pipelines, a workflow builder, and full self-host. If you want to prototype a RAG product without writing the retrieval layer yourself, this is where most serious operators start.

[Skill] coleam00/excalidraw-diagram-skill, Gives Claude Code (and any coding agent) the ability to generate Excalidraw diagrams from a prompt. 4,070 stars. Useful for architecture planning inside a Claude session without leaving the terminal.

[Skill] elementalsouls/Claude-OSINT, Two paired Claude skills with 90+ recon modules, 48 secret-regex patterns, 80+ dorks, 9 read-only credential validators, and 27 attack-path templates. 1,929 stars. If you are doing any security auditing with Claude, this is a significant shortcut.

[Tool] simonw/llm, Simon Willison's CLI and Python library for running prompts against any model. Supports local models via plugin, chains cleanly into shell scripts. The non-obvious move: pipe it into a cron job and you have a scheduled prompt runner without a single line of orchestration code.

[Tool] paul-gauthier/aider, Terminal-based AI pair programmer, git-aware, supports multi-file edits and architect mode. The thing most people miss: architect mode splits the reasoning step from the edit step, which cuts bad rewrites significantly. Worth running alongside Claude Code for complex refactors.

[Tool] yamadashy/repomix, Packs a repo into a single file for LLM context. Outputs XML, Markdown, or plain text. The specific use: pass the whole codebase to a model in one shot for a code review, a security audit, or seeding a new agent's context. Faster than any file-by-file approach.

- Apple sues OpenAI for trade secret theft. The lawsuit alleges OpenAI used proprietary Apple materials. For operators building on top of either ecosystem, this is the lawsuit most likely to reshape API terms in the next 12 months. Watch it. (Washington Post)

- Meta pulled its Instagram image-scraping AI tool after public backlash. The feature automatically accessed public Instagram images to generate AI content. It is off now. The relevant signal: if your pipeline ingests public social content for training or generation, the legal and PR ground under that practice is actively shifting. (Washington Post)

- Bun rewrote a major component in Rust in 11 days with AI assistance, for $165K in tokens. The Pragmatic Engineer broke down what that pace means for small teams doing rewrites they have been deferring for years. The cost number is real. If you have a backlog of "we'll rewrite that properly someday" items, that someday just got a price tag. (Pragmatic Engineer)

- Seedream 5.0 Pro is now available on Vercel's AI Gateway. Text-to-image with accurate text rendering, available via the gateway today. If your pipeline needs image generation with legible text, this is worth a test call before defaulting to your current provider. (Vercel Changelog)

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 alarm that could sleep through its own fire.

The component whose job is to tell you things broke is itself a thing that can break. Almost nobody reviews it that way.

We run a watchdog over one of our automated loops. It went through code review and passed. It went through a second round and passed. The third round asked a different question, not "does this work" but "walk me through the exact failure it exists to catch, line by line, and show me the alert leaving the building." It could not. Under the precise condition it was built to detect, the watchdog died before its alert went out. It watched for a failure that would also kill the messenger.

Three rounds. The first two reviewed the happy path of the unhappy path, which is the easiest trap in reviewing any monitoring code: you check that it detects the problem, and forget to check that detection survives the problem.

The reusable question is one sentence: in the exact scenario this alert exists for, does anything the alert depends on also fail? Run it against every monitor you own. Shared process, shared host, shared network, shared credentials, shared disk. Any yes means your alarm and your fire share a room.

Five questions we now ask of anything that watches anything: Who watches this when it fails? Does its alert path share a dependency with what it monitors? Does it fail loud or fail silent? When did it last actually fire, and would we notice if it never did again? Can we trigger a test failure on purpose, on a schedule?

That last one is the keeper. A monitor that has never been proven to fire is a decoration.

Have a watchdog you have never seen bark? Hit reply and tell us what it guards. We read every one.

Nine drops, three stack picks, four signals. The theme today was discipline: agents that block the bad commands, tools that trim the token waste, and a harness that handles the plumbing you have been writing by hand. The community is doing the safety work the labs are slow to ship. Might as well use it.

See you Tuesday.